JP7534869B2 - Accident prediction method, computer program, and accident prediction device - Google Patents

Description

The present invention relates to an accident prediction method, a computer program , and an accident prediction device for predicting the probability of a traffic accident occurring.

With technological developments in communication and IoT (Internet of Things), various information is collected as big data, and various technologies that utilize big data have been proposed. For example, technology is being developed that uses traffic information obtained from equipment installed on roads to predict the occurrence of accidents on roads.

Patent Document 1 proposes an accident prediction system that predicts accidents using first past traffic data measured in the past at each point on a route and second past traffic data measured at adjacent points upstream and downstream of the measurement points. In this system, the first past traffic data and the second past traffic data are associated with each other and learned to create an accident prediction table, and first current traffic data and second current traffic data measured in the present are acquired, and the likelihood of an accident occurring at each point is predicted using the accident prediction table and the first current traffic data and the second current traffic data.

JP 2017-84268 A

However, the accident prediction system described in Patent Document 1 has the problem that it can only predict the occurrence of current accidents based on current traffic data.

The present invention has been made in view of the above circumstances, and an object of the present invention is to provide an accident prediction method, a computer program , and an accident prediction device for predicting the probability of a future traffic accident occurring.

In one embodiment of the accident prediction method, machine learning is performed in advance using learning data in which time-series traffic information for a specified section of a road is associated with time-series accident information related to the presence or absence of an accident in the specified section, and an information processing device equipped with a learning model that accepts the time-series traffic information as input and outputs the accident information acquires the traffic information, inputs the time-series traffic information from the acquired traffic information before a specified time point into the learning model, acquires the accident information output by the learning model, and predicts the probability of an accident occurring in the specified section after the specified time point based on the acquired accident information.

In one embodiment, it is expected that the probability of a future traffic accident occurring can be predicted.

1 is a schematic diagram for explaining an overview of an accident prediction system according to an embodiment of the present invention; 1 is a block diagram showing a configuration of an accident prediction server device according to an embodiment of the present invention; 2 is a block diagram showing a configuration of a traffic server device according to the present embodiment. FIG. FIG. 2 is a schematic diagram showing an example of the configuration of a traffic DB. 2 is a block diagram showing a configuration of a terminal device according to the present embodiment; FIG. 2 is a schematic diagram for explaining a road on which the accident prediction model according to the present embodiment predicts the probability of a traffic accident occurring. FIG. 3 is a schematic diagram for explaining learning data used in the learning process of the accident prediction model according to the present embodiment. FIG. FIG. 2 is a schematic diagram showing a configuration example of an accident prediction model according to the present embodiment. 5 is a flowchart showing a procedure of a learning process of an accident prediction model performed by the accident prediction server device according to the present embodiment. 5 is a flowchart showing a procedure of an accident prediction process performed by the accident prediction server device according to the present embodiment. 10 is a schematic diagram showing an example of a display of an accident prediction result by a terminal device according to the present embodiment. FIG. 10 is a schematic diagram showing an example of a display of an accident prediction result by a terminal device according to the present embodiment. FIG. 10 is a schematic diagram showing an example of a display of an accident prediction result by a terminal device according to the present embodiment. FIG. 10 is a schematic diagram showing an example of a display of an accident prediction result by a terminal device according to the present embodiment. FIG. FIG. 13 is a schematic diagram for explaining learning data used in the learning process of the accident prediction model according to the modified example.

Specific examples of an accident prediction system according to an embodiment of the present invention will be described below with reference to the drawings. Note that the present invention is not limited to these examples, but is intended to include all modifications within the scope of the claims and meaning equivalent to the claims.

<System Overview>
1 is a schematic diagram for explaining an outline of an accident prediction system according to the present embodiment. The accident prediction system according to the present embodiment is a system in which a server system 1 composed of one or more server devices predicts the probability of future traffic accidents occurring on a road based on traffic information obtained from a plurality of traffic sensors 5 installed on the road and a vehicle 8, etc. The server system 1 transmits the predicted result of the probability of traffic accidents occurring as accident prediction information to, for example, a terminal device 6 used by a traffic controller, a terminal device 7 such as a smartphone used by a user, a vehicle 8, or a roadside device 9 installed on the road, etc.

The server system 1 is configured to include one or more server devices, and these server devices work together to realize the process of predicting the probability of future traffic accidents. In this embodiment, the server system 1 is configured with three server devices: an accident prediction server device 2, a traffic server device 3, and a weather server device 4, but this is just one example, and the number of server devices that make up the server system 1 and their division of roles, etc., can be changed as appropriate.

The traffic sensors 5 are placed, for example, at predetermined intervals of several hundred meters on the road, and detect the traffic volume, traffic density, vehicle speed, occupancy, etc. on the road, and transmit the detection results as traffic information to the traffic server device 3. Note that occupancy refers to the ratio of the space occupied by vehicles on the road at a certain point in time, or the ratio of time that vehicles are present at a certain point during a certain period of time. The traffic sensor 5 may be, for example, a device that obtains this traffic information by analyzing images taken of the road by a camera, a so-called traffic counter. The traffic sensor 5 transmits traffic information to the traffic server device 3 at intervals of, for example, several to several tens of seconds.

Vehicle 8 transmits information such as its own speed, acceleration, or position information, so-called probe data, as traffic information to traffic server device 3. Vehicle 8 can obtain this information from an on-board device such as a car navigation device, a drive recorder, a digital tachograph, or an ETC (Electronic Toll Collection) on-board device, or from a terminal device such as a smartphone carried by a user of vehicle 8 and an application program installed on this terminal device. Vehicle 8 transmits traffic information to traffic server device 3 via wireless communication such as a mobile phone communication network. Traffic server device 3 receives traffic information from multiple vehicles 8 traveling on the road.

The traffic server device 3 is a device that stores and accumulates various traffic information obtained from the traffic sensors 5 and vehicles 8, etc. in a database. In this embodiment, the traffic server device 3 acquires traffic information from the traffic sensors 5 and vehicles 8, but traffic information may be acquired via various devices other than the traffic sensors 5 and vehicles 8. Although not shown in the figure, the traffic server device 3 is provided with information on traffic accidents that have occurred on this road as accident information from, for example, a company that manages the road or a police station. The traffic server device 3 stores traffic information related to the road as well as accident information that has occurred on this road in the database.

The weather server device 4 provides meteorological information, such as weather, precipitation, accumulated precipitation, air temperature, road surface temperature, wind direction, sunrise time, or sunset time, to the accident prediction server device 2. The weather information of the weather server device 4 includes information on past weather measurements and information on predicted future weather. The weather server device 4 may be provided by, for example, the Japan Meteorological Agency or a weather forecasting company, and does not have to be included in the server system 1 of the accident prediction system. In any case, the accident prediction server device 2 can obtain past measured weather information and future weather forecast information as meteorological information.

The accident prediction server device 2 acquires traffic information and accident information acquired by the traffic server device 3 and stored in a database, and weather information provided by the weather server device 4, and performs processing to predict the probability of future traffic accidents occurring on the road based on the acquired information. As will be described in detail later, the accident prediction server device 2 generates a learning model that predicts the probability of future traffic accidents occurring by performing machine learning in advance using past time-series traffic information, accident information, and weather information as teacher data. The generated learning model outputs the probability of future traffic accidents occurring in response to input of time-series traffic information and weather information. The accident prediction server device 2 acquires time-series traffic information and weather information within a predetermined period including the latest time point (current time) from the traffic server device 3 and the weather server device 4 for the generated learning model, inputs the traffic accident occurrence probability output by the learning model, and predicts the probability of future traffic accidents occurring on this road.

The prediction results of the accident prediction server device 2 are given to various devices such as terminal devices 6, 7, vehicles 8, or roadside devices 9, and these devices display the accident prediction information to provide information or warn users such as road controllers or drivers of vehicles 8. Based on the accident prediction information displayed on the terminal device 6, road controllers can, for example, strengthen patrols on roads with a high accident probability. The terminal device 7 can, for example, display a warning about roads with a high accident probability in an app (application program) that displays a map, to alert users. Similarly, the vehicle 8 can, for example, display a warning on a map displayed by a car navigation device to alert the driver. The roadside device 9 is, for example, a display device such as an electronic bulletin board installed on a road, and can display information about the probability of a traffic accident to vehicles 8 traveling on the road to alert the driver.

<Device Configuration>
2 is a block diagram showing the configuration of the accident prediction server device 2 according to the present embodiment. The accident prediction server device 2 according to the present embodiment is configured to include a processing unit 21, a memory unit (storage) 22, a communication unit (transceiver) 23, and the like. The processing unit 21 is configured using an arithmetic processing device such as a central processing unit (CPU), a micro-processing unit (MPU) or a graphics processing unit (GPU), a read only memory (ROM), and a random access memory (RAM), and the like. The processing unit 21 reads and executes a server program 22a stored in the memory unit 22 to perform various processes such as a process of generating a learning model for predicting a traffic accident and a process of predicting a traffic accident using the generated learning model.

The storage unit 22 is configured using a large-capacity storage device such as a hard disk. The storage unit 22 stores various programs executed by the processing unit 21 and various data required for the processing of the processing unit 21. In this embodiment, the storage unit 22 stores a server program 22a executed by the processing unit 21 and an accident prediction model 22b that predicts the probability of a traffic accident occurring.

In this embodiment, the server program 22a is provided in a form recorded on a recording medium 99 such as a memory card or an optical disk, and the accident prediction server device 2 reads the server program 22a from the recording medium 99 and stores it in the storage unit 22. However, the server program 22a may be written in the storage unit 22, for example, during the manufacturing stage of the accident prediction server device 2. In addition, for example, the server program 22a may be distributed by another remote server device or the like and acquired by the accident prediction server device 2 through communication. For example, the server program 22a may be read from the recording medium 99 by a writing device and written to the storage unit 22 of the accident prediction server device 2. The server program 22a may be provided in a form distributed via a network, or may be provided in a form recorded on the recording medium 99.

The accident prediction model 22b provided in the accident prediction server device 2 according to this embodiment is a machine learning model that receives input of time-series traffic information and weather information, and outputs information related to the probability of future traffic accidents. The machine learning model performs a predetermined calculation on the input value and outputs the calculation result, and data such as coefficients and thresholds of functions that define this calculation are stored in the memory unit 32 as the accident prediction model 22b. In this embodiment, the accident prediction model 22b may be various models such as a recurrent neural network (RNN), a long short-term memory (LSTM), a gated recurrent unit (GRU), a multi-layer perceptron (MLP), or a convolutional neural network (CNN). Note that these models are existing technologies, so detailed explanations are omitted.

In the present embodiment, the accident prediction model 22b is pre-trained by the accident prediction server device 2. The learning process is a process of setting appropriate values for the coefficients and thresholds of each neuron constituting the neural network using a large amount of pre-given learning data. The accident prediction model 22b according to the present embodiment is trained by a so-called supervised machine learning method using learning data (teacher data) in which traffic information and accident information stored in a database by the traffic server device 3 are associated with weather information obtained from the weather server device 4. The accident prediction model 22b may be re-trained periodically, for example, once a month or once a year. In the present embodiment, the accident prediction server device 2 performs the learning process, but this is not limited thereto. A device other than the accident prediction server device 2 may perform the learning process, and the accident prediction server device 2 may acquire the trained accident prediction model 22b from this device and store it in the storage unit 32.

In addition, in this embodiment, the accident prediction server device 2 generates multiple accident prediction models 22b and stores the multiple accident prediction models 22b in the storage unit 22. In more detail, the accident prediction server device 2 generates an accident prediction model 22b for each road to be predicted, and generates an accident prediction model 22b for each type of accident to be predicted. Therefore, for example, if there are three roads to be predicted and there are two types of accidents, six accident prediction models 22b are generated. When a road to be predicted is specified, the accident prediction server device 2 reads out from the storage unit 32 and uses the accident prediction models 22b for all accident types related to the specified road, and predicts the probability of traffic accidents occurring for all types.

The communication unit 23 communicates with various devices via a network N including the Internet, a mobile phone communication network, and a LAN (Local Area Network), etc. In this embodiment, the communication unit 23 communicates with devices such as the traffic server device 3, the weather server device 4, the terminal devices 6 and 7, the vehicle 8, and the road device 9 via the network N. The communication unit 23 transmits data provided by the processing unit 21 to other devices, and provides data received from other devices to the processing unit 21.

The storage unit 22 may be an external storage device connected to the accident prediction server device 2. The accident prediction server device 2 may be a multi-computer including multiple computers, or may be a virtual machine virtually constructed by software. The accident prediction server device 2 is not limited to the above configuration, and may include, for example, a reading unit that reads information stored in a portable storage medium, an input unit that accepts operational input, or a display unit that displays images.

In addition, in the processing unit 21 of the accident prediction server device 2 according to this embodiment, the processing unit 21 reads out and executes the server program 22a stored in the memory unit 22, thereby realizing an information acquisition unit 21a, a prediction processing unit 21b, an information provision unit 21c, a learning processing unit 21d, and the like as software functional units. Note that these functional units are functional units related to the processing of accident prediction, machine learning, and the like, and illustrations and descriptions of other functional units are omitted.

The information acquisition unit 21a performs a process of acquiring information necessary for the accident prediction server device 2 to perform traffic accident prediction processing and accident prediction model 22b learning processing. The information acquisition unit 21a acquires traffic information and accident information from the traffic server device 3 by communicating with the traffic server device 3 through the communication unit 23. The information acquisition unit 21a also acquires weather information from the weather server device 4 by communicating with the weather server device 4 through the communication unit 23. When the accident prediction server device 2 performs learning processing, the information acquisition unit 21a acquires traffic information and accident information for a predetermined period (for example, several months to several years) used as learning data from the traffic server device 3, and acquires weather information from the weather server device 4. When the accident prediction server device 2 performs prediction processing, the information acquisition unit 21a acquires the latest traffic information from the traffic server device 3 at a frequency of, for example, once every few minutes, and acquires weather information from the weather server device 4 at a frequency of, for example, once every hour. The weather information acquired at this time includes information on the latest weather observed at that time and predicted future weather.

The prediction processing unit 21b performs processing to predict the probability of a traffic accident using the traffic information and weather information acquired by the information acquisition unit 21a and the accident prediction model 22b stored in the storage unit 22. The prediction processing unit 21b inputs time-series traffic information and weather information for a predetermined period including the current time (latest time) to the accident prediction model 22b. The accident prediction model 22b outputs the probability of a traffic accident occurring in the future after a predetermined time has elapsed from the current time for this input time-series information. The prediction processing unit 21b acquires the information output by the accident prediction model 22b and uses it as a prediction result of the probability of a traffic accident occurring.

The information providing unit 21c performs a process of providing information related to traffic accidents by transmitting the result of the prediction of the probability of occurrence of a traffic accident by the prediction processing unit 21b to another device via the communication unit 23. For example, the information providing unit 21c may only transmit the numerical value of the probability of occurrence of a traffic accident predicted by the prediction processing unit 21b to the other device as data. In this case, the other device generates a display screen and a message for presenting the probability of occurrence of a traffic accident to the user based on the received data. For example, the information providing unit 21c may generate data for the other device to display a screen that displays the probability of occurrence of a traffic accident in a graph, a screen that colors roads on a map according to the probability of occurrence of a traffic accident, or a screen that displays a list of roads with a high probability of occurrence of a traffic accident, based on the probability of occurrence of a traffic accident predicted by the prediction processing unit 21b, and transmit this display data to the other device. The information providing unit 21c may provide information on the probability of occurrence of a traffic accident in any data format.

The learning processing unit 21d performs a process of generating the accident prediction model 22b by machine learning. Based on the traffic information and accident information acquired from the traffic server device 3 and the weather information acquired from the weather server device 4, the learning processing unit 21d generates learning data (teacher data) in which these pieces of information correspond in time series. The learning processing unit 21d performs machine learning using the generated learning data for a learning model such as RNN or LSTM prepared in advance. The learning processing unit 21d stores the generated learned accident prediction model 22b in the storage unit 22. The learning processing unit 21d may also perform a re-learning process on the learned accident prediction model 22b at intervals of, for example, one month or one year. In this case, the learning processing unit 21d can generate learning data based on traffic information, accident information, and weather information accumulated after the time when the learning process was previously performed, and perform the re-learning process.

Figure 3 is a block diagram showing the configuration of the traffic server device 3 according to this embodiment. The traffic server device 3 according to this embodiment is configured with a processing unit 31, a memory unit (storage) 32, a communication unit (transceiver) 33, etc. The processing unit 31 is configured using an arithmetic processing device such as a CPU, MPU or GPU, a ROM, a RAM, etc. The processing unit 31 reads and executes a server program 32a stored in the memory unit 32, thereby performing various processes such as a process of collecting information from the traffic sensors 5 and vehicles 8, etc., and a process of providing the collected information to the accident prediction server device 2, etc.

The storage unit 32 is configured using a large-capacity storage device such as a hard disk. The storage unit 32 stores various programs executed by the processing unit 31 and various data required for the processing of the processing unit 31. In this embodiment, the storage unit 32 stores a server program 32a executed by the processing unit 31. The storage unit 32 also includes a traffic DB (database) 32b for storing and accumulating information collected from the traffic sensor 5 and the vehicle 8, etc.

In this embodiment, the server program 32a is provided in a form recorded on a recording medium 98 such as a memory card or an optical disk, and the traffic server device 3 reads the server program 32a from the recording medium 98 and stores it in the storage unit 32. However, the server program 32a may be written to the storage unit 32, for example, during the manufacturing stage of the traffic server device 3. In addition, for example, the server program 32a may be distributed by another remote server device or the like and acquired by the traffic server device 3 through communication. For example, the server program 32a may be read from the recording medium 98 by a writing device and written to the storage unit 32 of the traffic server device 3. The server program 32a may be provided in a form distributed via a network, or may be provided in a form recorded on the recording medium 99.

Figure 4 is a schematic diagram showing an example of the configuration of the traffic DB 32b. In the traffic DB 32b according to this embodiment, information such as "date and time" and "location" is stored in association with "traffic information" and "accident information". "Date and time" is information indicating the date and time when the corresponding traffic information was measured by the traffic sensor 5 and the vehicle 8. "Location" is information indicating the location where the corresponding traffic information was measured. In the illustrated example, the location is shown as a combination of identification information attached to the road and identification information attached to the traffic sensor 5, such as "Road A Sensor A01" and "Road A Sensor A02". However, the "location" may be information such as latitude and longitude, or other information, as long as it can uniquely identify a location on the road.

The "traffic information" stored in the traffic DB 32b may include information such as "speed", "traffic volume" and "occupancy rate". "Speed" is, for example, the average speed per unit time of vehicles traveling on a road. "Traffic volume" is, for example, the number of vehicles passing a road per unit time. "Occupancy rate" is, for example, the proportion of time that vehicles were on the road in a given time. Note that the traffic information shown in the figure is an example and is not limited thereto, and various information other than speed, traffic volume and occupancy rate may be stored as traffic information in the traffic DB 32b. In this embodiment, the "traffic information" stored in the traffic DB 32b is based on information acquired from the traffic sensor 5. However, the "traffic information" stored in the traffic DB 32b may include information based on information acquired from the vehicle 8. The traffic server device 3 may store the information acquired from the traffic sensor 5 itself as "traffic information", or may store information that has been calculated or processed on the acquired information. For example, the traffic server device 3 can acquire information from multiple vehicles 8 traveling at points on a target road, and calculate information such as speed, traffic volume, or occupancy rate from the acquired information, or information that can be substituted for these, and store this in "speed" in the traffic DB 32b.

The "accident information" stored in the traffic DB 32b may include information such as "accident occurrence" and "accident type". "Accident occurrence" is information indicating whether or not a traffic accident has occurred at this location, and either a value of "0 (no accident)" or "1 (accident)" is stored. "Accident type" is information indicating the type of traffic accident that has occurred. In this embodiment, traffic accidents are classified into two types of "accident types", namely, "vehicle contact accidents" which are contacts or rear-end collisions between vehicles, and "facility contact accidents" which are contacts or collisions between vehicles and facilities other than vehicles. However, the "accident types" are not limited to these two types, and various accident types may be adopted. For example, the "accident types" may include contact accidents between people and vehicles, or accidents due to natural disasters such as falling rocks. "Accident information" may be based on information acquired from, for example, the traffic sensor 5 and the vehicle 8, or may be based on accident information obtained from, for example, a road management company or a police station. The traffic server device 3 may, for example, obtain traffic accident information in real time and store it in the traffic DB 32b, or may obtain traffic accident information after the fact. In addition, the "accident information" may be information input by an administrator or the like who manages the traffic server device 3.

The communication unit 33 communicates with various devices via a network N including the Internet, a mobile phone communication network, a LAN, and the like. In this embodiment, the communication unit 33 communicates with devices such as the accident prediction server device 2, the vehicle 8, and the road device 9 via the network N. The communication unit 33 transmits data provided by the processing unit 31 to other devices, and provides data received from other devices to the processing unit 31.

The storage unit 32 may be an external storage device connected to the traffic server device 3. The traffic DB 32b may be provided in a device separate from the traffic server device 3, in which case the traffic server device 3 transmits the acquired traffic information and accident information to the separate device. The traffic server device 3 may be a multi-computer including multiple computers, or may be a virtual machine virtually constructed by software. The traffic server device 3 is not limited to the above configuration, and may include, for example, a reading unit that reads information stored in a portable storage medium, an input unit that accepts operational input, or a display unit that displays images.

In addition, the processing unit 31 of the traffic server device 3 according to this embodiment has an information acquisition unit 31a and an information provision unit 31b, etc., implemented as software functional units by the processing unit 31 reading and executing the server program 32a stored in the storage unit 32. The information acquisition unit 31a performs processing to acquire traffic information and accident information by communicating with the traffic sensor 5 and the vehicle 8, etc., through the communication unit 33. For example, the information acquisition unit 31a acquires traffic information from the traffic sensor 5 at a frequency of about once every few seconds to several tens of seconds. For example, the information acquisition unit 31a acquires traffic information by receiving traffic information transmitted from the vehicle 8 through the communication unit 33. For example, the information acquisition unit 31a acquires accident information by accepting input of accident information by an administrator, etc. The information acquisition unit 31a stores the acquired traffic information and accident information in the traffic DB 32b of the storage unit 32 after performing appropriate calculations or processing, etc., as necessary.

The information providing unit 31b performs a process of transmitting the traffic information and accident information stored in the traffic DB 32b to the accident prediction server device 2 by communication via the communication unit 33. For example, the information providing unit 31b reads out and transmits traffic information and accident information for a predetermined period required for the accident prediction server device 2 to perform a learning process from the traffic DB 32b. The information providing unit 31b also communicates with the accident prediction server device 2 at a frequency of, for example, once every few minutes, and periodically transmits traffic information including the latest information obtained from the traffic sensor 5 and the vehicle 8. Note that the traffic server device 3 may transmit information to the accident prediction server device 2 on its own initiative, or may transmit the information in response to a request from the accident prediction server device 2.

Fig. 5 is a block diagram showing the configuration of the terminal device 6 according to this embodiment. The terminal device 6 according to this embodiment is configured to include a processing unit 61, a memory unit (storage) 62, a communication unit (transceiver) 63, a display unit (display) 64, and an operation unit 65. The terminal device 6 may be configured using an information processing device such as a general-purpose personal computer or a tablet-type terminal device.

The processing unit 61 is configured using an arithmetic processing device such as a CPU or an MPU, a ROM, a RAM, etc. The processing unit 61 reads and executes a program 62a stored in the storage unit 62 to perform various processes such as displaying traffic accident prediction information.

The storage unit 62 is configured using, for example, a magnetic storage device such as a hard disk or a non-volatile memory element such as a flash memory. The storage unit 62 stores various programs executed by the processing unit 61 and various data necessary for the processing of the processing unit 61. In this embodiment, the storage unit 62 stores the program 62a executed by the processing unit 61. In this embodiment, the program 62a is distributed by a remote server device or the like, and the terminal device 6 acquires it through communication and stores it in the storage unit 62. However, the program 62a may be written to the storage unit 62, for example, during the manufacturing stage of the terminal device 6. For example, the program 62a may be read by the terminal device 6 from the program 62a recorded on a recording medium 97 such as a memory card or an optical disk and stored in the storage unit 62. For example, the program 62a may be read by a writing device from the program 62a recorded on the recording medium 97 and written to the storage unit 62 of the terminal device 6. The program 62a may be provided in the form of distribution via a network, or may be provided in the form of being recorded on the recording medium 97.

The communication unit 63 communicates with various devices via a network N including an in-house LAN, a wireless LAN, the Internet, etc. In this embodiment, the communication unit 63 communicates with the accident prediction server device 2 via the network N. The communication unit 63 transmits data provided by the processing unit 61 to other devices, and provides data received from other devices to the processing unit 61.

The display unit 64 is configured using a liquid crystal display or the like, and displays various images, characters, etc. based on the processing of the processing unit 61. The operation unit 65 accepts user operations and notifies the processing unit 61 of the accepted operations. For example, the operation unit 65 accepts user operations through an input device such as a mechanical button or a touch panel provided on the surface of the display unit 64. Furthermore, for example, the operation unit 65 may be an input device such as a mouse and a keyboard, and these input devices may be configured to be removable from the terminal device 6.

In addition, in the terminal device 6 according to this embodiment, the processing unit 61 reads and executes the program 62a stored in the storage unit 62, whereby the display processing unit 61a and the like are realized in the processing unit 61 as software-like functional units. Note that the program 62a may be a program dedicated to the accident prediction system according to this embodiment, or may be a general-purpose program such as an internet browser or web browser.

The display processing unit 61a performs processing to display various characters, images, and the like on the display unit 64. In this embodiment, the display processing unit 61a receives information on the predicted results of the probability of occurrence of a traffic accident transmitted from the accident prediction server device 2 via the communication unit 63, and performs a screen display on the display unit 64 to inform the user of the probability of occurrence of a traffic accident based on the received information. The display processing unit 61a can perform various displays, such as displaying a graph showing the change over time in the probability of occurrence of an accident, or displaying a road map color-coded according to the probability of occurrence of an accident.

<Generation process of accident prediction model>
FIG. 6 is a schematic diagram for explaining roads for which the accident prediction model 22b according to the present embodiment predicts the probability of occurrence of a traffic accident. In this figure, a road that does not have a branch or a merger over a certain distance, such as an expressway or a toll road, is shown as an example. These roads have branch points and merge points such as junctions and interchanges. In this figure, the branch points or merge points of the roads are shown with solid circles. Also, in this figure, the traffic sensor 5 installed on the road is shown with a dashed circle. The roads include two roads, that is, a road on which a vehicle travels from one branch point to the other branch point, and a road on which a vehicle travels from the other branch point to the one branch point, that is, so-called uphill and downhill roads, but in this embodiment, the uphill and downhill roads are considered to be different roads. Note that the prediction by the accident prediction model 22b is not limited to expressways, and roads such as general roads may be predicted.

In the accident prediction system according to this embodiment, specific sections of a road are determined in advance as target sections for accident prediction, and an accident prediction model 22b is generated for each target section. The target sections for accident prediction are determined in advance, for example, by an administrator of the system. In this figure, the target sections are shown surrounded by dashed lines.

The accident prediction model 22b according to this embodiment predicts the probability of an accident occurring in a target section of a road based on traffic information relating to the target section and the surrounding sections of the target section. The surrounding sections are sections of roads within a specified range of the target section and connect to the target section. In this figure, the surrounding sections are shown surrounded by a two-dot chain line. The surrounding sections may be determined by an administrator of the system, as with the target section, or may be determined by the accident prediction server device 2, etc. for the target section.

The traffic information used to predict the probability of an accident may include traffic information obtained from traffic sensors 5 installed in the target section and surrounding sections, information such as speed, traffic volume or occupancy rate calculated from information obtained from one or more vehicles 8 traveling in these sections, information that can be substituted for these, or information that is a combination of these pieces of information.

The traffic server device 3 acquires traffic information such as the speed and traffic volume of vehicles traveling on the target section of the road from traffic sensors 5 installed in the target section and surrounding sections of the road, and stores the information in the traffic DB 32b. The traffic server device 3 also acquires traffic information such as position information and speed from one or more vehicles 8 that are present in the target section of the road and surrounding sections, and stores the information in the traffic DB 32b.

In this embodiment, one accident prediction model 22b is generated by machine learning for each target section of the road (and for each uphill and downhill section of the target section). One accident prediction model 22b is generated for each type of traffic accident, and in this embodiment, there are two types of traffic accidents: facility contact accidents and vehicle contact accidents. Therefore, in order to predict the probability of traffic accidents occurring for multiple target sections, multiple accident prediction models 22b must be generated in advance, and the number of accident prediction models 22b is (number of target sections) × (number of types of traffic accidents) = (number of accident prediction models 22b). Note that the above types of traffic accidents are examples and are not limited to these, and various types such as personal injury accidents may be included. The number of types of accidents is not limited to two, and may be one type or three or more types.

Figure 7 is a schematic diagram for explaining the learning data used in the learning process of the accident prediction model 22b according to this embodiment. In this embodiment, the accident prediction server device 2, which performs the learning process of the accident prediction model 22b, generates learning data to be used in the learning process of the accident prediction model 22b based on information that can be acquired from the traffic server device 3 and the weather server device 4. The accident prediction server device 2 acquires three types of traffic information, namely, speed information, traffic volume information, and occupancy information, which are stored and accumulated in the traffic DB 32b by the traffic server device 3. In the illustrated example, for each of the speed information, traffic volume information, and occupancy information, information measured by the traffic sensor 5 at n points P1 to Pn is used. In addition, these multiple pieces of information are arranged in a time series, for example, every 5 minutes. For each piece of information every 5 minutes, if the measurement period of the traffic sensor 5 is 5 minutes, for example, the measured value in each period can be adopted, and if the measurement period is shorter than 5 minutes, a representative value such as the average value or maximum value of multiple measured values can be adopted, and if the measurement period is longer than 5 minutes, for example, one measured value can be adopted multiple times. In this example, the information is divided every 5 minutes, but the time interval for dividing the information does not have to be 5 minutes and can be set appropriately based on the measurement period, amount of information, etc.

The accident prediction server device 2 also associates meteorological information such as precipitation and cumulative precipitation that can be obtained from the weather server device 4 with the time-series traffic information every 5 minutes. The amount of precipitation is, for example, the amount of rain that fell in 5 minutes, and the cumulative precipitation is, for example, the amount of rain that fell in 2 hours. The amount of precipitation and cumulative precipitation do not necessarily have to be the actual measured value of the amount of rain that fell on the road in the target section, but may be the amount of precipitation observed in the area such as a city, town, or village that includes the target section. The weather server device 4 acquires measured values such as precipitation from each area where a measuring instrument or the like is installed, and stores and accumulates them in a database. The accident prediction server device 2 acquires meteorological information for the area corresponding to the target section of the road from the meteorological information stored in the weather server device 4, and associates the acquired meteorological information with the traffic information. Note that in this example, the amount of precipitation and cumulative precipitation are used, but various other meteorological information such as weather, temperature, road surface temperature, wind speed, sunrise time, or sunset time may also be used.

The accident prediction server device 2 also acquires information on the presence or absence of traffic accidents stored in the traffic DB 32b by the traffic server device 3. In this embodiment, the presence or absence of a traffic accident is binary information of 1 (accident) or 0 (no accident). The traffic DB 32b stores the types of accidents in association with the presence or absence of an accident, and the accident prediction server device 2 only needs to acquire information on the presence or absence of an accident related to the accident type targeted by the accident prediction model 22b to be created. The accident prediction server device 2 creates the data shown in FIG. 7, which associates traffic information, weather information, and the presence or absence of an accident in chronological order. This data created by the accident prediction server device 2 may be, for example, a quantity spanning a period of several months to several years. The accident prediction server device 2 generates learning data to be used in the learning process of the accident prediction model 22b by extracting desired information from this data.

The accident prediction server device 2 extracts traffic information from the traffic server device 3, which includes the reference time and is included in the range from a past time point that is a predetermined time (e.g., 2 hours) before the reference time point to the reference time point. The accident prediction server device 2 also extracts weather information from the weather server device 4, which includes the past time point that is a predetermined time before the reference time point to a future time point after a predetermined time (e.g., 30 minutes, 1 hour, or 2 hours) has passed from the reference time point. The accident prediction server device 2 also extracts information on the presence or absence of an accident at a future time point after a predetermined time (e.g., 30 minutes, 1 hour, or 2 hours) has passed from the reference time point. The accident prediction server device 2 sets data in which the extracted traffic information from the past time point to the reference time point, the extracted weather information from the past time point to the future time point, and the extracted presence or absence of an accident at the future time point are associated with each other as learning data (information surrounded by a thick line in FIG. 6). The accident prediction server device 2 generates multiple learning data by, for example, changing the reference time point by 5 minutes and acquiring similar information.

Figure 8 is a schematic diagram showing an example of the configuration of the accident prediction model 22b according to this embodiment. In this embodiment, the accident prediction model 22b employs an LSTM learning model. The LSTM is a type of RNN suitable for handling time-series information, and is a learning model capable of learning the long-term dependencies of time-series information. The LSTM is constructed by combining multiple cells called LSTM cells (or LSTM blocks), each of which has a memory that holds the cell state and one or more gates that control the input and output of information to this memory. Note that the internal structure of the LSTM cell is not illustrated or described in detail because it is an existing technology, but any structure may be adopted.

The accident prediction model 22b according to this embodiment is configured with ten LSTM cells A to J, as shown in FIG. 8. Four of the five LSTM cells A to E in the first stage, LSTM cells A to D, receive traffic and weather information included in the learning data from two hours ago to the reference time, divided into four parts for every 30 minutes. That is, information from two hours before the reference time to one hour and 30 minutes before is input to LSTM cell A, information from one hour and 30 minutes before to one hour before is input to LSTM cell B, information from one hour before to 30 minutes before is input to LSTM cell C, and information from 30 minutes before to the reference time is input to LSTM cell D. Weather information from the reference time to a future time is input to LSTM cell E. Also, output information from LSTM cells A to D, which receive information from the previous time, is input to LSTM cells B to E.

The five LSTM cells F-J in the latter stage receive output information from the corresponding LSTM cells A-E in the previous stage. LSTM cells G-J receive output information from LSTM cells F-I, which receive information corresponding to the previous time. Ultimately, the output information from LSTM cell J becomes the prediction result of traffic accidents on the road in the target section. In this embodiment, the output of LSTM cell J is normalized by a softmax function to a number in the range of 0 to 1, and this number is treated as the probability of a traffic accident occurring. In the learning process, this prediction result is associated with the presence or absence of an accident in the learning data.

Note that the configuration of the accident prediction model 22b shown in FIG. 8 is an example and is not limited to this. The number of LSTM cells in the accident prediction model 22b does not have to be 10, but may be 9 or less or 11 or more. The arrangement and connection of multiple LSTM cells are not limited to those shown in the figure. The accident prediction model 22b may also adopt a learning model such as GRU or MLP instead of the LSTM learning model.

The accident prediction server device 2 performs a learning process for the accident prediction model 22b shown in FIG. 8 using the learning data shown in FIG. 7. The learning process by the accident prediction server device 2 is performed by a method such as the gradient descent method, the stochastic gradient descent method, or the backpropagation method. These learning methods are existing technologies, so a detailed description is omitted.

Figure 9 is a flowchart showing the procedure of the learning process of the accident prediction model 22b performed by the accident prediction server device 2 according to this embodiment. The processing unit 21 of the accident prediction server device 2 according to this embodiment accepts input, for example, from the administrator of the system, to determine the target section of the road to be the subject of the learning process (step S1) and to determine the surrounding sections (step S2).

The information acquisition unit 21a of the processing unit 21 acquires traffic information and accident information related to the target section determined in step S1 and the surrounding sections determined in step S2 from the traffic server device 3 (step S3). The information acquisition unit 21a also acquires meteorological information related to the target section determined in step S1 and the surrounding sections determined in step S2 from the weather server device 4 (step S4).

The learning processing unit 21d of the processing unit 21 generates learning data that associates the traffic information, weather information, and accident information acquired in steps S3 and S4 (step S4). At this time, the learning processing unit 21d generates learning data that associates time-series traffic information from a reference time point to a past time point a predetermined time before, weather information from the past time point to a future time point, and accident information for a future time point a predetermined time after the reference time point. The learning processing unit 21d performs machine learning using the learning data generated in step S5 on the accident prediction model 22b in the initial state that was generated in advance (step S6). The learning processing unit 21d stores the learned accident prediction model 22b that was subjected to machine learning in step S6 in the storage unit 22 (step S7), and ends the process.

<Accident prediction processing>
The accident prediction server device 2 according to this embodiment communicates with the traffic server device 3, for example, once every five minutes, and acquires traffic information such as speed, traffic volume, and occupancy rate stored in the traffic DB 32b by the traffic server device 3. The traffic information acquired by the accident prediction server device 2 from the traffic server device 3 includes information stored in the traffic DB 32b after the previously acquired traffic information, up to the latest information. The accident prediction server device 2 stores the traffic information acquired from the traffic server device 3 in the storage unit 22 for at least a period required for inputting the information into the accident prediction model 22b, for example, two hours.

The accident prediction server device 2 also communicates with the weather server device 4, for example, once every five minutes, to acquire weather information such as precipitation and accumulated precipitation. The weather information acquired by the accident prediction server device 2 at this time includes past weather information from a past time point a predetermined time ago (for example, two hours) to the present time, and future weather information from the present time to a future time point (prediction time point) a predetermined time later. The past weather information may be actual measured values such as precipitation and accumulated precipitation, or may be predicted values predicted by the weather server device 4. For example, the accident prediction server device 2 may acquire weather information for one hour of the past two hours as actual measured values and the other hour as predicted values. The future weather information is a predicted value. The accident prediction server device 2 stores the weather information acquired from the weather server device 4 in the memory unit 22 for at least the period required for inputting the weather information to the accident prediction model 22b.

The accident prediction server device 2 predicts the probability of an accident occurring by inputting the traffic information acquired from the traffic server device 3 and the weather information acquired from the weather server device 4 into the learned accident prediction model 22b stored in the storage unit 22. The accident prediction server device 2 reads out one corresponding accident prediction model 22b from the storage unit 22 according to the road section and type of accident to be predicted, and inputs time-series traffic information from the past to the present and weather information from the past to the future to this accident prediction model 22b. The accident prediction model 22b outputs a prediction result of the probability of a traffic accident occurring at a future time after a predetermined time has elapsed from the present time according to the input traffic information and weather information. The accident prediction server device 2 acquires the prediction result output by the accident prediction model 22b and transmits it to, for example, a terminal device 6, etc., thereby displaying the prediction result of the probability of a traffic accident occurring on the terminal device 6, etc.

The accident prediction server device 2 may read out the accident prediction model 22b corresponding to each type of accident for the road section to be predicted, and execute the above-mentioned accident prediction for each type of accident, thereby making accident predictions for all types of accidents. The accident prediction server device 2 may read out the accident prediction model 22b corresponding to each target road section, and execute the above-mentioned accident prediction for each type of accident, thereby making accident predictions for all predictable sections. The accident prediction server device 2 can predict the accident occurrence probability for all accident factors for all predictable sections of the road, and transmit the prediction results to the terminal device 6.

Figure 10 is a flowchart showing the procedure of the accident prediction process performed by the accident prediction server device 2 according to this embodiment. The prediction processing unit 21b of the processing unit 21 of the accident prediction server device 2 according to this embodiment determines the target section for predicting the probability of occurrence of a traffic accident (step S21). The information acquisition unit 21a of the processing unit 21 acquires traffic information related to the target section determined in step S21 and its surrounding sections from the traffic server device 3 (step S22). At this time, the information acquisition unit 21a acquires traffic information in time series from a past time point a predetermined time ago to the present time point (or a time point to be used as the basis for prediction). The information acquisition unit 21a also acquires weather information related to the target section determined in step S21 and its surrounding sections from the weather server device 4 (step S23). At this time, the information acquisition unit 21a acquires weather information in time series from the present time point (or a time point to be used as the basis for prediction) to a future time point after a predetermined time.

The prediction processing unit 21b reads out from the storage unit 22 an accident prediction model 22b corresponding to one accident type from among the accident prediction models 22b corresponding to the target section determined in step S21 (step S24). The prediction processing unit 21b inputs the time-series traffic information and weather information acquired in steps S22 and S23 to the accident prediction model 22b read out in step S24 (step S25). The prediction processing unit 21b acquires the prediction result output by the accident prediction model 22b in response to the information input in step S25 (step S26).

The prediction processing unit 21b determines whether prediction results for all accident types have been obtained for the target section (step S27). If prediction results for all accident types have not been obtained (S27: NO), the prediction processing unit 21b returns to step S24 and performs the same processing using the accident prediction model 22b corresponding to another accident type.

When prediction results have been obtained for all accident types (S27: YES), the prediction processing unit 21b determines whether prediction results have been obtained for all sections of the predictable road (step S28). When prediction results have not been obtained for all sections (S28: NO), the prediction processing unit 21b returns to step S21, determines another section as the target section, and performs similar processing. When prediction results have been obtained for all sections (S28: YES), the information providing unit 21c of the processing unit 21 transmits these prediction results to the terminal devices 6, 7, the vehicle 8, the road device 9, etc. via the communication unit 33 (step S29), and ends the processing.

The terminal device 6 that receives the prediction results from the accident prediction server device 2 performs processing to display the predicted results of the probability of traffic accidents occurring in the target section of the road. Simply displaying the probability of traffic accidents occurring as the prediction result of the accident prediction server device 2 makes it difficult for a user who sees this display to determine whether or not an accident is more likely to occur than usual. Therefore, the accident prediction system according to this embodiment displays information such as the ratio of the predicted value to the actual value of the accident occurrence probability so that the user can easily determine whether or not an accident is more likely to occur than usual. Below, several examples of displays by the accident prediction system according to this embodiment are explained.

Fig. 11 is a schematic diagram showing an example of the display of accident prediction results by the terminal device 6 according to this embodiment. The terminal device 6 according to this embodiment displays, for example, the change in the accident occurrence probability per day for the section of the road to be predicted, in the graph shown in Fig. 11. The graph in this example shows the accident occurrence rate for vehicle contact accidents on the "ABC Expressway Upbound", with the horizontal axis representing the time from 0:00 to 24:00 and the vertical axis representing the accident occurrence probability.

The terminal device 6 displays, as a line graph, the "average accident probability" for each time based on, for example, the past year's actual results (shown by a solid line in FIG. 11) and the "average AI predicted value" (shown by a dashed line in FIG. 11), which is the average value of the accident probability for each time predicted by the accident prediction server device 2 over the past year. The terminal device 6 also displays the accident probability for each time predicted by the accident prediction server device 2 for today (or the target prediction day) as the "AI predicted value," and displays a point on the graph corresponding to the predicted value (shown by a triangle in FIG. 11). These displays allow controllers and others using the terminal device 6 to compare and confirm the past traffic accident actual values, the average value of past accident predictions by AI, and today's predicted value.

Furthermore, the terminal device 6 according to this embodiment displays a graph with the horizontal axis representing the time from 0:00 to 24:00 and the vertical axis representing the magnification of the AI prediction result, superimposed on the graph of the accident occurrence rate. That is, the terminal device 6 displays a point on the graph corresponding to the magnification (indicated by "X" in FIG. 11) on which the above-mentioned "AI predicted value" corresponds to the "average accident occurrence probability" as the "magnification of the AI predicted value to the average accident occurrence probability." This display allows air traffic controllers and others using the terminal device 6 to easily recognize how high or low the AI predicted value of the accident occurrence probability for today is compared to the past actual value. That is, the terminal device 6 can not only display a simple accident occurrence probability, but also display whether or not an accident is more likely to occur than usual based on the "magnification of the AI predicted value to the average accident occurrence probability."

The terminal device 6 also displays a point on the graph (indicated by a "+" in FIG. 11) that corresponds to the "multiple of the average AI predicted value" to indicate how many times the "AI predicted value" is compared to the "average AI predicted value." These displays allow air traffic controllers and others using the terminal device 6 to easily recognize how high or low the AI predicted value for today's accident probability is compared to the average value of past predictions. In other words, the terminal device 6 does not simply show the accident probability, but also takes into account the prediction tendency of the AI (accident prediction model 22b) and can display whether an accident is more likely to occur than usual using the "multiple of the average AI predicted value."

In order for the terminal device 6 to display these values, information such as past actual values and past predicted values is required. In this embodiment, the information necessary for these displays is transmitted by the accident prediction server device 2 to the terminal device 6. The accident prediction server device 2 may store the information necessary for display by itself, or may obtain the necessary information from the traffic server device 3. For example, the accident prediction server device 2 may store the prediction results so far as history in the storage unit 22, calculate the average value for the terminal device 6 to display as the "average of AI predicted values", and transmit it to the terminal device 6. In addition, for example, the accident prediction server device 2 may obtain information on the actual occurrence or absence of accidents over the past year from the traffic server device 3, tally up the number of accidents that occurred for each hour, and transmit the ratio of the number of days that accidents occurred to one year (365 days) as the probability to be displayed as the "average accident occurrence probability" to the terminal device 6. In addition, the calculation of the "average accident occurrence probability" may be performed by the traffic server device 3, and the accident prediction server device 2 may obtain the calculation result from the traffic server device 3. The calculation method for the "average AI predicted value" and "average accident occurrence probability" is not limited to the method described above, and various calculation methods can be used.

12 is a schematic diagram showing an example of a display of an accident prediction result by the terminal device 6 according to the present embodiment. The terminal device 6 according to the present embodiment displays, for example, the accident occurrence probability at a specific time for a section of a road to be predicted, in the graph shown in FIG. 12. The graph in this example corresponds to information for a specific time cut out from the graph shown in FIG. 11 and displayed as a bar graph. The terminal device 6 displays three items, "average accident occurrence probability," "average AI predicted value," and "AI predicted value," as shown in FIG. 11, side by side as bar graphs. Furthermore, the terminal device 6 displays a point indicating the magnification of the AI predicted result, superimposed on the bar graph of the accident occurrence probability. The terminal device 6 displays a point corresponding to the "multiplication of the average accident occurrence probability of the AI predicted value" (indicated by "X" in FIG. 12) and a point corresponding to the "multiplication of the average AI predicted value" (indicated by "+" in FIG. 12) superimposed on the bar graph. These displays allow a controller using the terminal device 6 to check the accident occurrence probability by focusing on a specific time.

FIG. 13 is a schematic diagram showing an example of a display of accident prediction results by the terminal device 6 according to this embodiment. The terminal device 6 according to this embodiment displays an overall map of roads for which accident prediction is possible below the title character string "AI prediction value (today)", and also displays arrows indicating the accident prediction probability in different colors superimposed on this map. The direction of each arrow indicates whether the road is going up or down. These displays allow the controller or the like using the terminal device 6 to easily confirm which parts of the entire road have a high probability of causing a traffic accident. Note that the map images necessary for the terminal device 6 to perform these displays are stored in the terminal device 6, but may be transmitted to the terminal device 6 by the accident prediction server device 2.

Figure 14 is a schematic diagram showing an example of a display of accident prediction results by the terminal device 6 according to this embodiment. The display example shown in Figure 14 is an example in which the "multiplication factor of the average accident occurrence probability of the AI predicted value" or the "multiplication factor of the average value of the AI predicted value" is displayed in color on a map in the same display method as in Figure 13. The terminal device 6 according to this embodiment displays an overall road map below the title string of "multiplication factor of the AI predicted value" and superimposes an arrow on this map indicating the multiplication factor of the AI predicted value in color.

For example, on the "AI predicted value (today)" screen shown in FIG. 13, a road with an accident probability of over 30% is located in the upper left corner of the map. However, on the "AI predicted value multiplier" screen shown in FIG. 14, it can be seen that the multiplier for this road is a low multiplier of 1.0 to 1.25. In other words, by comparing the screens of FIG. 13 and FIG. 14, a road controller or the like can understand that this road is one that has a high accident probability on a regular basis, and that the accident probability is not high only today. This allows a controller or the like using the terminal device 6 to easily confirm which locations have a high AI accident probability prediction value compared to past actual values or predicted values, etc.

In this embodiment, three examples are shown in which the terminal device 6 displays the predicted accident probability results, but the terminal device 6 may display the predicted results using a display method other than these. The terminal device 7 and the vehicle 8 can also perform displays similar to those of the terminal device 6. As for the roadside device 9, for example, a warning message can be displayed to the driver of the vehicle 8 on the roadside device 9 installed on a road where the accident probability predicted by the accident prediction server device 2 exceeds a threshold value.

<Summary>
In the accident prediction system according to the present embodiment, machine learning of the accident prediction model 22b is performed in advance using learning data in which time-series traffic information and accident information related to a target section of a road are associated, and the accident prediction server device 2 predicts the accident occurrence probability in the target section using the learned accident prediction model 22b. The accident prediction model 22b accepts time-series traffic information as an input and outputs accident information. The accident prediction server device 2 acquires traffic information from the traffic server device 3, and inputs time-series traffic information before a predetermined time to the accident prediction model 22b. The accident prediction server device 2 acquires accident information output by the accident prediction model 22b in response to this input, and predicts the accident occurrence probability in the target section of the road at a future time point after the predetermined time point based on the acquired accident information. As a result, the accident prediction system according to the present embodiment is expected to accurately predict the occurrence probability of future traffic accidents.

The accident prediction system according to this embodiment generates an accident prediction model 22b for each type of accident. The accident prediction server device 2 stores multiple accident prediction models 22b generated for each type of accident in the memory unit 22, and predicts the accident occurrence probability for each type of accident. Since the conditions under which an accident is likely to occur vary depending on the type of accident, by preparing an accident prediction model 22b for each type of accident, the accident prediction server device 2 is expected to predict the occurrence probability of a traffic accident with greater accuracy. However, the accident prediction model 22b does not have to be generated separately for each type of accident, and may be a model that predicts the occurrence probability regardless of the type of accident. The accident prediction model 22b may also be generated as a single model that outputs multiple accident occurrence probabilities.

The accident prediction system according to this embodiment also performs a learning process for the accident prediction model 22b using learning data in which weather information is associated with traffic information and accident information. This allows predictions to be made taking into account weather information as well as traffic information, so the accident prediction server device 2 is expected to predict the probability of traffic accidents with greater accuracy. Furthermore, when predicting the probability of traffic accidents, the accident prediction server device 2 may obtain weather information predicted by the weather server device 4 and input it to the accident prediction model 22b. This allows the accident prediction server device 2 to perform a process for predicting the probability of accidents based on predicted weather information even when actual measured values of weather information cannot be obtained, so it is expected to be able to make accident predictions quickly.

In addition, in the accident prediction system according to this embodiment, the accident prediction server device 2 predicts the probability of an accident occurring in a target section of the road to be predicted based on traffic information related to the target section and the sections surrounding this target section. As a result, the accident prediction server device 2 is expected to make more accurate predictions compared to when predictions are made based on traffic information only on the target section.

In the accident prediction system according to the present embodiment, the terminal device 6 etc. outputs (displays) the predicted value of the accident probability for the target section of the road along with the actual value (average accident probability) based on the information provided by the accident prediction server device 2. The terminal device 6 etc. also outputs the magnification of the predicted value to the actual value as information indicating the relationship between the actual value and the predicted value. The accident prediction system according to the present embodiment also stores the accident probability predicted by the accident prediction server device 2, and outputs the average value of the past accident probability (average of the AI predicted value) stored by the terminal device 6 etc. along with the predicted value of the accident probability. The terminal device 6 etc. also outputs the magnification of the predicted value to the average value. By the terminal device 6 etc. outputting this information, the user using the terminal device 6 etc. can be expected to easily and reliably grasp the prediction result of the accident prediction server device 2.

(Modification)
In the above embodiment, the accident prediction model 22b is machine-learned using learning data in which traffic information, weather information, and accident information are associated, but this is not limited to the above. The accident prediction model 22b according to the modified example is machine-learned using learning data in which calendar information is associated in addition to traffic information, weather information, and accident information. The accident prediction model 22b that has been machine-learned accepts input of time-series traffic information, weather information, and calendar information, and outputs the probability of occurrence of a traffic accident. The calendar information is, for example, information such as date, day of the week, time, and holiday, and in this embodiment, is information that can be stored and managed by the accident prediction server device 2.

Figure 15 is a schematic diagram for explaining learning data used in the learning process of the accident prediction model 22b according to the modified example. The accident prediction server device 2 according to the modified example associates traffic information such as speed, traffic volume, and occupancy rate acquired from the traffic server device 3, weather information such as precipitation and accumulated precipitation acquired from the weather server device 4, calendar information such as date, day of the week, and time managed by the accident prediction server device 2, and accident information such as the presence or absence of a traffic accident, in a time series. The accident prediction server device 2 extracts traffic information and calendar information included in the range from a past time point including a reference time point and a predetermined time (e.g., 2 hours) before the reference time point to the reference time point from the traffic information and calendar information associated with the time series. The accident prediction server device 2 also extracts weather information included in the range from a past time point a predetermined time before the reference time point to a future time point a predetermined time (e.g., 30 minutes, 1 hour, or 2 hours) after the reference time point from the weather information associated with the time series. The accident prediction server device 2 also extracts information on the presence or absence of an accident at a future time point after a predetermined time (e.g., 30 minutes, 1 hour, or 2 hours) has elapsed from the reference time point. The accident prediction server device 2 sets as learning data (information surrounded by a thick line in FIG. 15) data that associates the extracted traffic information and calendar information from the reference time point up to a predetermined time ago, the extracted weather information from the past to the future, and the extracted presence or absence of an accident at a future time point. The accident prediction server device 2 generates multiple pieces of learning data by, for example, changing the reference time point by 5 minutes and acquiring similar information.

The accident prediction model 22b according to the modified example receives as input time-series traffic information and calendar information from a past time point a predetermined time before the present time point (reference time point) to the present time point, and weather information from past time points to a future time point, and outputs the probability of a traffic accident occurring at a future time point after the reference time point. The accident prediction server device 2 performs a learning process for the accident prediction model 22b using the learning data shown in FIG. 15.

The accident prediction server device 2 according to the modified example predicts the probability of an accident by inputting the traffic information acquired from the traffic server device 3, the actual and predicted values of the weather information acquired from the weather server device 4, and the calendar information managed by the accident prediction server device 2 itself into the learned accident prediction model 22b. At this time, the accident prediction server device 2 acquires the date, day of the week, time, etc. of the time of prediction as calendar information based on the date, day of the week, time, etc. managed by its own clock function and calendar function, and inputs the acquired calendar information into the accident prediction model 22b. The accident prediction model 22b outputs the probability of a future traffic accident occurring a predetermined time after the present time according to the input traffic information, weather information, and calendar information. The accident prediction server device 2 acquires the prediction result output by the accident prediction model 22b and transmits it to, for example, a terminal device 6, etc., thereby displaying the predicted result of the probability of a traffic accident occurring on the terminal device 6, etc.

In the accident prediction system according to the modified example of the above configuration, the accident prediction server device 2 performs machine learning of the accident prediction model 22b using learning data in which traffic information, weather information, calendar information, and accident information are associated. The accident prediction server device 2 inputs the traffic information acquired from the traffic server device 3, the weather information acquired from the weather server device 4, and calendar information acquired by its own calendar function or the like to the learned accident prediction model 22b, and acquires the occurrence probability of a traffic accident output by the accident prediction model 22b. As a result, if there is a tendency for the occurrence of traffic accidents according to the calendar, such as traffic accidents being more likely to occur on Fridays or in December, the accident prediction model 22b can predict the occurrence probability of a traffic accident taking this tendency into account, and it is expected that the accuracy of the prediction can be improved.

The embodiments disclosed herein are illustrative in all respects and should not be considered limiting. The scope of the present invention is indicated by the claims, not by the meaning described above, and is intended to include all modifications within the scope and meaning equivalent to the claims.

1 Server system 2 Accident prediction server device (accident prediction device)
3 Traffic server device 4 Weather server device 5 Traffic sensor 6 Terminal device 7 Terminal device 8 Vehicle 9 Road device 21 Processing unit 21a Information acquisition unit 21b Prediction processing unit 21c Information provision unit 21d Learning processing unit 22 Memory unit 22a Server program (computer program)
22b Accident prediction model (learning model)
23 Communication unit 31 Processing unit 31a Information acquisition unit 31b Information provision unit 32 Storage unit 32a Server program 32b Traffic DB
33 Communication unit 61 Processing unit 61a Display processing unit 62 Storage unit 62a Program 63 Communication unit 64 Display unit 65 Operation unit 97 to 99 Recording medium

Claims (14)

an information processing device that has a learning model that performs machine learning in advance using learning data in which time-series traffic information relating to a predetermined section of a road is associated with time-series accident information relating to the presence or absence of an accident in the predetermined section, receives the time-series traffic information as an input, and outputs the accident information,
Acquire the traffic information;
From the acquired traffic information, traffic information in a time series prior to a predetermined time point is input to the learning model;
Acquire the accident information output by the learning model;
predicting an accident occurrence probability in the predetermined section after the predetermined time based on the acquired accident information;
A past accident occurrence probability is acquired as a performance value related to the occurrence of an accident in the predetermined section,
outputting the acquired actual value together with the predicted value of the accident occurrence probability;
Accident prediction methods. an information processing device that has a learning model that performs machine learning in advance using learning data in which time-series traffic information relating to a predetermined section of a road is associated with time-series accident information relating to the presence or absence of an accident in the predetermined section, receives the time-series traffic information as an input, and outputs the accident information,
Acquire the traffic information;
From the acquired traffic information, traffic information in a time series prior to a predetermined time point is input to the learning model;
Acquire the accident information output by the learning model;
predicting an accident occurrence probability in the predetermined section after the predetermined time based on the acquired accident information;
A past accident occurrence probability is acquired as a performance value related to the occurrence of an accident in the predetermined section,
outputting information indicating a correlation between the acquired actual value and the predicted value together with the predicted value of the accident occurrence probability;
Accident prediction methods. an information processing device that has a learning model that performs machine learning in advance using learning data in which time-series traffic information relating to a predetermined section of a road is associated with time-series accident information relating to the presence or absence of an accident in the predetermined section, receives the time-series traffic information as an input, and outputs the accident information,
Acquire the traffic information;
From the acquired traffic information, traffic information in a time series prior to a predetermined time point is input to the learning model;
Acquire the accident information output by the learning model;
predicting an accident occurrence probability in the predetermined section after the predetermined time based on the acquired accident information;
storing the predicted accident occurrence probability;
outputting, together with the predicted value of the accident occurrence probability, a magnification of the predicted value with respect to an average value of the stored plurality of the accident occurrence probabilities;
Accident prediction methods. The information processing device includes a learning model for each of a plurality of accident types that has been subjected to machine learning using the learning data prepared for each accident type,
The information processing device,
inputting the traffic information in time series before a predetermined time point into each of the plurality of learning models;
Acquire the accident information output from each of the plurality of learning models;
predicting an accident occurrence probability for each accident type in the specified section after the specified time point based on the acquired plurality of pieces of accident information;
The accident prediction method according to any one of claims 1 to 3. the information processing device includes a learning model that is machine-learned using learning data in which the traffic information and the accident information are associated with time-series weather information,
The information processing device,
Obtaining the weather information in chronological order;
inputting the traffic information and the weather information in time series before a predetermined time point into the learning model;
The accident prediction method according to any one of claims 1 to 4. The information processing device includes a learning model that is machine-learned using learning data in which the traffic information and the accident information are associated with calendar information,
The information processing device,
Obtaining the calendar information;
inputting the traffic information in time series before a predetermined time point and the calendar information at the predetermined time point into the learning model;
The accident prediction method according to any one of claims 1 to 5. The traffic information includes information related to roads in the specified section and surrounding roads in the specified section.
The accident prediction method according to any one of claims 1 to 6. storing the predicted accident occurrence probability;
outputting an average value of the stored plurality of accident occurrence probabilities together with the predicted value of the accident occurrence probability;
The accident prediction method according to any one of claims 1 to 7. On the computer,
Machine learning is performed in advance using learning data in which time-series traffic information for a specific section of a road is associated with time-series accident information for the specific section, and a learning model is used to receive the time-series traffic information as an input and output the accident information.
The traffic information is acquired in a time series manner;
From the acquired traffic information, traffic information in a time series prior to a predetermined time point is input to the learning model;
Acquire the accident information output by the learning model;
predicting an accident occurrence probability in the predetermined section after the predetermined time based on the acquired accident information;
A past accident occurrence probability is acquired as a performance value related to the occurrence of an accident in the predetermined section,
outputting the acquired actual value together with the predicted value of the accident occurrence probability;
A computer program that executes a process. On the computer,
Machine learning is performed in advance using learning data in which time-series traffic information for a specific section of a road is associated with time-series accident information for the specific section, and a learning model is used to receive the time-series traffic information as an input and output the accident information.
The traffic information is acquired in a time series manner;
From the acquired traffic information, traffic information in a time series prior to a predetermined time point is input to the learning model;
Acquire the accident information output by the learning model;
predicting an accident occurrence probability in the predetermined section after the predetermined time based on the acquired accident information;
A past accident occurrence probability is acquired as a performance value related to the occurrence of an accident in the predetermined section,
outputting information indicating a correlation between the acquired actual value and the predicted value together with the predicted value of the accident occurrence probability;
A computer program that executes a process. On the computer,
A learning model that performs machine learning in advance using learning data in which time-series traffic information for a specific section of a road is associated with time-series accident information for the specific section, receives the time-series traffic information as an input, and outputs the accident information,
The traffic information is acquired in a time series manner;
From the acquired traffic information, traffic information in a time series prior to a predetermined time point is input to the learning model;
Acquire the accident information output by the learning model;
predicting an accident occurrence probability in the predetermined section after the predetermined time based on the acquired accident information;
storing the predicted accident occurrence probability;
outputting, together with the predicted value of the accident occurrence probability, a magnification of the predicted value with respect to an average value of the stored plurality of the accident occurrence probabilities;
A computer program that executes a process. a learning model that performs machine learning in advance using learning data in which time-series traffic information for a predetermined section of a road is associated with time-series accident information for the presence or absence of an accident in the predetermined section, receives the time-series traffic information as an input, and outputs the accident information;
A traffic information acquisition unit that acquires the traffic information in time series;
a prediction unit that inputs time-series traffic information before a predetermined time point from the traffic information acquired by the traffic information acquisition unit into the learning model, acquires the accident information output by the learning model, and predicts an accident occurrence probability in the predetermined section after the predetermined time point;
a performance value acquiring unit that acquires a past accident occurrence probability as a performance value related to the occurrence of an accident in the predetermined section;
and an output unit that outputs the acquired actual value together with the predicted value of the accident occurrence probability. a learning model that performs machine learning in advance using learning data in which time-series traffic information for a predetermined section of a road is associated with time-series accident information for the presence or absence of an accident in the predetermined section, receives the time-series traffic information as an input, and outputs the accident information;
A traffic information acquisition unit that acquires the traffic information in time series;
a prediction unit that inputs time-series traffic information before a predetermined time point from the traffic information acquired by the traffic information acquisition unit into the learning model, acquires the accident information output by the learning model, and predicts an accident occurrence probability in the predetermined section after the predetermined time point;
a performance value acquiring unit that acquires a past accident occurrence probability as a performance value related to the occurrence of an accident in the predetermined section;
and an output unit that outputs information indicating a correlation between the acquired actual value and the predicted value together with the predicted value of the accident occurrence probability. a learning model that performs machine learning in advance using learning data in which time-series traffic information for a predetermined section of a road is associated with time-series accident information for the presence or absence of an accident in the predetermined section, receives the time-series traffic information as an input, and outputs the accident information;
A traffic information acquisition unit that acquires the traffic information in time series;
a prediction unit that inputs time-series traffic information before a predetermined time point from the traffic information acquired by the traffic information acquisition unit to the learning model, acquires the accident information output by the learning model, and predicts an accident occurrence probability in the predetermined section after the predetermined time point;
A storage unit that stores the predicted accident occurrence probability;
and an output unit that outputs, together with the predicted value of the accident occurrence probability, a magnification of the predicted value with respect to an average value of the stored plurality of the accident occurrence probabilities.

Images