JP7633005B2 - TRAFFIC CONDITION DETERMINATION DEVICE, TRAFFIC CONDITION DETERMINATION METHOD, COMPUTER PROGRAM, AND LEARNING MODEL - Google Patents

Description

This disclosure relates to a traffic condition determination device, a traffic condition determination method, a computer program, and a learning model.

Patent Document 1 discloses a detection device that detects abnormal traffic flow (traffic flow with reduced traffic capacity due to a sudden event such as an accident). The detection device disclosed in Patent Document 1 predicts the traffic state quantity a predetermined time from the present time based on traffic state quantity data up to the present, sets upper and lower thresholds for determining abnormal traffic flow based on the predicted value of the traffic state quantity, determines whether the actual traffic state quantity detected after the predetermined time is outside the range between the upper and lower thresholds, and determines that the traffic flow is abnormal if it is outside the range.

Japanese Patent Application Publication No. 3-209599

In the detection device disclosed in Patent Document 1, if it is not possible to accurately predict the traffic state quantity a predetermined time from the present time, it is not possible to appropriately set a threshold value for determining abnormal traffic flow, and the accuracy of detecting abnormal traffic flow decreases. Furthermore, traffic flow after the occurrence of an unexpected event changes depending on the content and severity of the unexpected event such as an accident, the traffic volume at the time the unexpected event occurs, the road shape where the unexpected event occurs, etc. It is difficult to set an appropriate threshold value in accordance with the various circumstances related to the unexpected event as described above.

A traffic condition determination device according to one aspect of the present disclosure includes a learning model that generates output data indicating traffic conditions at a specific point from input data indicating the traffic conditions at the specific point, and a condition determination unit that determines whether or not an abnormal traffic condition has occurred based on the input data and the output data, and the learning model is an autoencoder that includes an input layer, one or more intermediate layers, and an output layer.

A traffic condition determination method according to one aspect of the present disclosure includes a step of generating output data indicating traffic conditions at a specific point from input data indicating the traffic conditions at the specific point using a learning model, and a step of determining whether or not an abnormal traffic condition has occurred based on the input data and the output data, and the learning model is an autoencoder including an input layer, one or more intermediate layers, and an output layer.

A computer program according to one aspect of the present disclosure is a computer program for causing a computer to detect a traffic abnormality, and causes the computer to execute a step of generating output data indicating the traffic condition at a specific point from input data indicating the traffic condition at the specific point using a learning model, and a step of determining whether or not a traffic abnormality has occurred based on the input data and the output data, and the learning model is an autoencoder including an input layer, one or more intermediate layers, and an output layer.

A learning model according to one aspect of the present disclosure is a learning model configured by an autoencoder, and includes an input layer that accepts input data indicating traffic conditions at a specific point, an output layer that generates output data indicating traffic conditions at the specific point, and one or more intermediate layers provided between the input layer and the output layer.

The present disclosure can be realized not only as a traffic state determination device having the above-mentioned characteristic configuration, but also as a traffic state determination method having steps of characteristic processing in the traffic state determination device. The present disclosure can be realized as a computer program that causes a computer to function as a communication control device, or as a learning model for traffic state determination that is generated by machine learning executed by the traffic state determination device or another device. Furthermore, the present disclosure can be realized as a semiconductor integrated circuit as part or all of the traffic state determination device, or as a traffic state determination system including the traffic state determination device.

This disclosure makes it possible to accurately detect abnormal conditions in traffic.

1 is a diagram illustrating an example of an overall configuration of a traffic signal control system. 1 is a block diagram showing an example of a configuration of a traffic condition determination device according to an embodiment. FIG. 2 is a diagram illustrating types of learning models in the traffic condition determination device according to the embodiment. FIG. 2 is a schematic diagram showing an example of the configuration of a learning model according to an embodiment. FIG. 2 is a functional block diagram showing an example of functions of the traffic condition determination device according to the embodiment. 13 is a graph showing the time progression of reconstruction error at the same point. FIG. 2 is a diagram showing an example of a display screen of the display device according to the embodiment; FIG. 1 is a schematic diagram illustrating an example of machine learning according to an embodiment. 5 is a flowchart showing a procedure of a traffic condition determination process according to the embodiment. 10 is a flowchart illustrating a procedure of a learning process according to the embodiment.

Overview of the embodiments of the present disclosure
Below, an overview of the embodiments of the present disclosure will be listed and described.

(1) The traffic condition determination device according to the present embodiment includes a learning model that generates output data indicating the traffic condition at a specific point from input data indicating the traffic condition at the specific point, and a condition determination unit that determines whether or not a traffic abnormality occurs based on the input data and the output data, and the learning model is an autoencoder including an input layer, one or more intermediate layers, and an output layer. The autoencoder reduces the dimension of the input data (encodes it) and restores (decodes) it to obtain output data identical to the input data. When input data in a normal traffic condition is given to a learning model trained by various training data, output data that restores the input data with high accuracy is generated. Input data reflecting a traffic abnormality when an unexpected event occurs shows a tendency different from input data in a normal traffic condition. The learning model does not accurately restore the input data reflecting such a traffic abnormality, and generates output data different from the input data. Therefore, a traffic abnormality can be accurately detected based on the input data and the output data.

(2) The input data may include a congestion length, a traffic volume, and a travel time, the input layer may include a first input node that accepts the congestion length in the input data, a second input node that accepts the traffic volume in the input data, and a third input node that accepts the travel time in the input data, the output data may include a congestion length, a traffic volume, and a travel time, and the output layer may include a first output node that outputs the congestion length in the output data, a second output node that outputs the traffic volume in the output data, and a third output node that outputs the travel time in the output data. The traffic state can be accurately indicated by the congestion length, the traffic volume, and the travel time. Therefore, an abnormal traffic state can be accurately detected by using such input data.

(3) The traffic condition determination device according to this embodiment includes a plurality of the learning models corresponding to a plurality of time periods, and a classification unit that classifies traffic condition information indicating traffic conditions at the specific point according to the plurality of time periods, and the input data based on the traffic condition information classified by the classification unit may be input to the learning model corresponding to the same time period as the time period to which the input data corresponds. Traffic conditions tend to differ depending on the time period, such as morning and evening, night, and daytime. Therefore, abnormal traffic conditions can be accurately detected by providing input data classified according to the time period to a learning model created for each time period.

(4) The multiple learning models may correspond to multiple day types, and the classification unit may classify the traffic condition information according to the multiple day types. Traffic conditions tend to differ depending on the day type, such as weekdays, Saturdays, Sundays, and holidays. Therefore, abnormal traffic conditions can be accurately detected by providing input data classified according to the day type to the learning models created for each day type.

(5) The traffic condition determination device according to this embodiment may further include a normalization unit that normalizes traffic condition information indicating the traffic condition at the specific point and generates the input data. In this way, by normalizing the traffic condition information, inconsistencies between input data can be suppressed.

(6) The state determination unit may determine whether or not the abnormal traffic state has occurred by comparing the difference between the input data and the output data with a predetermined threshold. In a normal traffic state, the output data matches the input data well, whereas in an abnormal traffic state, the output data differs greatly from the input data. Therefore, by comparing the difference between the input data and the output data with a predetermined threshold, an abnormal traffic state can be easily and accurately detected.

(7) The state determination unit may determine that the abnormal traffic condition has occurred when the difference exceeds the threshold value a specified number of times in succession. When an unexpected incident occurs, an abnormal traffic condition such as a traffic jam continues. Such an abnormal traffic condition may not be resolved in a short period of time. Therefore, a continuing abnormal traffic condition can be detected by determining whether the difference exceeds a predetermined threshold value a specified number of times in succession.

(8) The traffic condition determination device according to this embodiment may further include a display device that displays a map that distinguishes between points where the traffic abnormality condition has occurred and points where the traffic abnormality condition has not occurred, based on the determination result of the condition determination unit. This allows the user to visually confirm the point where the traffic abnormality condition has occurred.

(9) The learning model may be generated by machine learning to output information that is similar to the training data when the training data indicating the traffic state is provided to the input layer. This makes it possible to create a learning model that can be used to detect abnormal traffic conditions.

(10) The traffic condition determination method according to this embodiment includes the steps of generating output data indicating traffic conditions at a specific point from input data indicating the traffic conditions at the specific point using a learning model, and determining whether or not an abnormal traffic condition has occurred based on the input data and the output data, where the learning model is an autoencoder including an input layer, one or more intermediate layers, and an output layer. The learning model accurately restores input data reflecting normal traffic conditions, but does not accurately restore input data reflecting abnormal traffic conditions. Therefore, an abnormal traffic condition can be accurately detected based on the input data and the output data.

(11) A computer program according to this embodiment is a computer program for causing a computer to detect an abnormal traffic state, and causes the computer to execute the steps of: generating output data indicating the traffic state at a specific point from input data indicating the traffic state at the specific point using a learning model; and determining whether or not an abnormal traffic state has occurred based on the input data and the output data. The learning model is an autoencoder including an input layer, one or more intermediate layers, and an output layer. The learning model accurately restores input data reflecting normal traffic states, but does not accurately restore input data reflecting abnormal traffic states. Therefore, an abnormal traffic state can be accurately detected based on the input data and the output data.

(12) The learning model according to this embodiment is a learning model configured by an autoencoder, and includes an input layer that accepts input data indicating traffic conditions at a specific point, an output layer that generates output data indicating traffic conditions at the specific point, and one or more intermediate layers provided between the input layer and the output layer. The learning model accurately restores input data that reflects normal traffic conditions, but does not accurately restore input data that reflects abnormal traffic conditions. Therefore, abnormal traffic conditions can be accurately detected based on the input data and output data.

<Details of the embodiment of the present disclosure>
Hereinafter, the details of the embodiments of the present invention will be described with reference to the drawings. Note that at least some of the embodiments described below may be combined in any desired manner.

[1. Traffic Signal Control System]
FIG. 1 is a diagram showing an example of the overall configuration of a traffic signal control system.
In FIG. 1 , a traffic signal control system 1 includes a plurality of signal lamps 2, a plurality of traffic signal controllers 4, a traffic information database 6 (hereinafter, the “database” may also be referred to as a “DB”), a roadside sensor 8, and a traffic condition determination device 10.

In FIG. 1, a signal lamp 2 and a traffic signal controller 4 are installed at each of a number of intersections Ci (i=1 to 12).
In FIG. 1, for the sake of simplicity, only one traffic light 2 is shown at each intersection Ci, but in reality, four traffic lights 2 are installed for the uphill and downhill directions of the intersecting roads.

The traffic information DB 6 and the traffic condition determination device 10 are installed in a traffic control center or the like.
The traffic signal controller 4 and the traffic information DB 6 are capable of wireless communication using a mobile communication system such as LTE (Long Term Evolution), and are capable of transmitting and receiving information to and from each other through wireless communication with each other.

The traffic signal controller 4 controls the light color (signal control) of the signal lamp 2.

The roadside sensor 8 is composed of, for example, a vehicle detector that ultrasonically detects vehicles 5 traveling directly below, a loop coil that detects vehicles 5 by inductance changes, or an image detector that processes camera images to measure traffic volume and vehicle speed.
Roadside sensors 8 are installed on the upstream side of each road extending from a target intersection Ci, and measure the number of vehicles entering the intersection Ci (incoming traffic volume), vehicle speed, travel time from the upstream intersection, and congestion length. Travel time is measured based on the time from when a vehicle is detected at the upstream intersection to when the vehicle is detected at the target intersection Ci. Congestion length is the length of the queue of vehicles that is congested. It is determined that a congestion has occurred when vehicles with vehicle speeds below a predetermined value are detected consecutively.

The roadside sensor 8 is connected to a traffic signal controller 4 installed at a target intersection Ci via a wired or wireless communication line. The roadside sensor 8 provides a measurement result to the traffic signal controller 4 via the communication line. The measurement result is used by the traffic signal controller 4 to control the light color of the signal lamp 2, and is also transferred by the traffic signal controller 4 to the traffic information DB 6.
The measurement results include the traffic volume, vehicle speed, travel time, etc. flowing into the intersection Ci. The roadside sensor 8 does not have to measure the travel time and congestion length. For example, a device (traffic information DB 6, etc.) installed in a traffic control center may obtain the detection time of a vehicle by each roadside sensor 8 and calculate the travel time. The device may obtain the vehicle speed detected by each roadside sensor 8, determine the congestion and calculate the congestion length.

The traffic information DB 6 and the traffic condition determination device 10 are connected to each other so that they can communicate with each other via a LAN (Local Area Network) or the like. The traffic information DB 6 stores traffic condition information indicating the traffic condition at each intersection Ci. The traffic condition information includes congestion length, traffic volume, and travel time.

The traffic condition determination device 10 detects abnormal traffic conditions based on the traffic condition information stored in the traffic information DB 6.

2. Configuration of the traffic condition determination device
FIG. 2 is a block diagram showing an example of the configuration of the traffic condition determination device 10 according to this embodiment.

The traffic condition determination device 10 includes a processor 101, a non-volatile memory 102, a volatile memory 103, and an input/output interface 104.

The volatile memory 103 is a volatile memory such as, for example, an SRAM (Static Random Access Memory) or a DRAM (Dynamic Random Access Memory). The non-volatile memory 102 is a non-volatile memory such as, for example, a flash memory, a hard disk, or a ROM (Read Only Memory). The non-volatile memory 102 stores a traffic condition determination program 105, which is a computer program, and data used to execute the traffic condition determination program 105. Each function of the traffic condition determination device 10 is realized by the processor 101 executing the traffic condition determination program 105. The traffic condition determination program 105 can be stored in a recording medium such as a flash memory, a ROM, or a CD-ROM.

The traffic condition determination program 105 includes a plurality of learning models 106. Furthermore, a training program 107 is stored in the non-volatile memory 102. The training program 107 is a computer program for machine learning, and is capable of constructing the learning models 106 by machine learning.

The processor 101 is, for example, a CPU (Central Processing Unit). However, the processor 101 is not limited to a CPU. The processor 101 may be a GPU (Graphics Processing Unit). In a specific example, the processor 101 is a multi-core GPU. The processor 101 may be, for example, a hardware logic circuit such as an ASIC (Application Specific Integrated Circuit), a gate array, or an FPGA (Field Programmable Gate Array). In this case, the hardware logic circuit is configured to be capable of executing processing similar to that of the traffic condition determination program 105.

The input/output interface 104 includes, for example, an Ethernet interface ("Ethernet" is a registered trademark), and is connected to the traffic information DB 6. Furthermore, the input/output interface 104 is connected to a display device 108 such as a liquid crystal display, and can output a video signal to the display device 108.

[3. Learning Model]
The learning model 106 is a computer program executable by the processor 101 and is used to detect traffic anomalies.

The traffic condition determination device 10 according to this embodiment stores a plurality of learning models 106. FIG. 3 is a diagram for explaining the types of learning models in the traffic condition determination device 10 according to this embodiment. Traffic conditions tend to differ depending on the time of day, such as morning and evening, night, and daytime. In order to respond to such trends in traffic conditions according to time of day, the learning model 106 is provided for each of a plurality of time periods. Specifically, a learning model 106 is provided corresponding to each time period of one hour. Traffic conditions tend to differ depending on the day type, such as weekdays, Saturdays, Sundays and holidays. In order to respond to such trends in traffic conditions according to day type, the learning model 106 is further provided for each of a plurality of day types. Specifically, a learning model 106 is provided corresponding to each day type, such as weekdays, Saturdays, and Sundays and holidays. In other words, one learning model 106 is provided corresponding to each time period on weekdays, one learning model 106 is provided corresponding to each time period on Saturdays, and one learning model 106 is provided corresponding to each time period on Sundays and holidays.

Figure 4 is a schematic diagram showing an example of the configuration of a learning model according to this embodiment. The learning model 106 is configured by an autoencoder, which is a type of neural network, and includes an input layer 106IN, an intermediate layer 106MID, and an output layer 106OUT.

The input layer 106IN includes three nodes N11, N12, and N13. Node N11 (first input node) accepts input of congestion length. Node N12 (second input node) accepts input of traffic volume. Node N13 (third input node) accepts input of travel time. Note that the number of nodes included in the input layer 106IN is not limited to three. For example, the input layer 106IN may include four or more nodes. In this case, nodes other than the three nodes that accept congestion length, traffic volume, and travel time may accept data related to traffic conditions other than congestion length, traffic volume, and travel time.

The intermediate layer 106MID is composed of one or more processing layers. In the example shown in FIG. 4, the intermediate layer 106MID has a one-layer structure. The intermediate layer 106MID includes one node N2. Note that the number of nodes included in the intermediate layer 106MID is not limited to one, but is less than the number of nodes included in the input layer 106IN.

The output layer 106OUT includes three nodes N31, N32, and N33. Node N31 (first output node) outputs the congestion length. Node N32 (second output node) outputs the traffic volume. Node N33 (third output node) outputs the travel time. The number of nodes included in the output layer 106OUT is not limited to three, but is the same as the number of nodes in the input layer 106IN. For example, the output layer 106OUT may include four or more nodes. In this case, nodes other than the three nodes that output the congestion length, traffic volume, and travel time may output data related to traffic conditions other than the congestion length, traffic volume, and travel time.

Each of nodes N31, N32, and N33 in the output layer 106OUT is connected to node N2 in the intermediate layer 106MID by an edge. Individual weights are set for the edges.

The data provided to each of the nodes N11, N12, and N13 in the input layer 106IN is weighted by the edge weight and passed to the node N2 in the intermediate layer 106MID. The node N2 in the intermediate layer 106MID executes the assigned processing on the provided data. The data processed by the node N2 in the intermediate layer 106MID is weighted by the edge weight and passed to each of the nodes N31, N32, and N33 in the output layer 106OUT. The data of each of the nodes N31, N32, and N33 included in the output layer 106OUT is the congestion length, traffic volume, and travel time. The weight of each edge is set so that the output data output from the nodes N31, N32, and N33 in the output layer 106OUT matches the input data provided to the nodes N11, N12, and N13 in the input layer 106IN.

The learning model 106 is constructed through a training process described below.

[4. Functions of the traffic condition determination device]
5 is a functional block diagram showing an example of functions of the traffic condition determination device 10 according to this embodiment. The traffic condition determination device 10 has the functions of an input unit 111, a classification unit 112, a normalization unit 113, a condition determination unit 114, and a video signal output unit 115.

Traffic information DB6 stores traffic condition information including congestion length, traffic volume, and travel time. The input unit 111 accepts the latest traffic condition information from the traffic information DB6. The traffic condition information includes, for example, the time when the information was obtained (measured by the roadside sensor 8).

The traffic condition information received by the input unit 111 is provided to the classification unit 112. The classification unit 112 classifies the provided traffic condition information according to the time period and the day type. For example, traffic condition information including time information of 8:45 on Friday is classified as day type: weekday and time period: 8:00 to 9:00, and traffic condition information including time information of 20:05 on Sunday is classified as day type: Sunday/holiday and time period: 20:00 to 21:00.

The traffic condition information classified by the classification unit 112 is provided to the normalization unit 113. The normalization unit 113 normalizes the congestion length, traffic volume, and travel time included in the traffic condition information. Specifically, the normalization unit 113 converts each of the congestion length, traffic volume, and travel time data into dimensionless quantities by dividing them by a representative value. This generates input data that are the dimensionless quantities of congestion length, traffic volume, and travel time.

The generated input data (traffic jam length, traffic volume, and travel time) is input to the learning model 106 corresponding to the day type and time period into which the traffic condition information (i.e., the input data) has been classified by the classification unit 112. For example, input data classified as day type: weekday and time period: 8:00 to 9:00 is provided to the learning model 106 corresponding to day type: weekday and time period: 8:00 to 9:00, and input data classified as day type: Sunday/holiday and time period: 20:00 to 21:00 is provided to the learning model 106 corresponding to day type: Sunday/holiday and time period: 20:00 to 21:00.

The learning model 106 compresses the dimensions of the input data and then restores them to generate output data. When input data indicating normal traffic conditions is input to the learning model 106, output data that roughly matches the input data is generated. When input data indicating abnormal traffic conditions is input to the learning model 106, output data that differs from the input data is generated.

The input data and output data of the learning model 106 are provided to the state determination unit 114. The state determination unit 114 determines whether or not an abnormal traffic condition has occurred based on the input data and the output data. Specifically, the state determination unit 114 determines whether or not an abnormal traffic condition has occurred by comparing the difference between the input data and the output data with a threshold value.

The function of the state determination unit 114 will be described in detail below. The state determination unit 114 calculates a reconstruction error according to the following formula: The reconstruction error is an example of a difference between input data and output data.
Reconstruction error = |Input congestion length - output congestion length| + |Input traffic volume - output traffic volume| + |Input travel time - output travel time|
However, the input congestion length is the congestion length included in the input data, the output congestion length is the congestion length included in the output data, the input traffic volume is the traffic volume included in the input data, the output traffic volume is the traffic volume included in the output data, the input travel time is the travel time included in the input data, and the output travel time is the travel time included in the output data.

The state determination unit 114 compares the calculated reconstruction error with a threshold value. The state determination unit 114 counts the number of times the reconstruction error continuously exceeds a predetermined threshold value (hereinafter referred to as the "continuous exceedance count"), and when the continuous exceedance count reaches a specified number of times, the traffic state is determined to be abnormal. In other words, the traffic state determination device 10 performs a series of processes for the same point, including acquisition of traffic state information, classification, normalization, generation of output data by a learning model, calculation of reconstruction error, and comparison of reconstruction error with a threshold value, multiple times. In these multiple processing cycles, when the continuous exceedance count is equal to or greater than the specified number of times, the state determination unit 114 determines that the traffic state is abnormal. When the continuous exceedance count is less than the specified number of times, the state determination unit 114 determines that the traffic state is normal. This makes it possible to prevent the traffic state from being erroneously determined to be abnormal when noise is included in the traffic state information.

Figure 6 is a graph showing the time transition of the reconstruction error at the same point. In Figure 6, the vertical axis indicates the reconstruction error, and the horizontal axis indicates the time. The solid line graph in the figure shows the time transition of the reconstruction error on November 10, and the dashed line graph shows the time transition of the reconstruction error on November 17. On November 10, no abnormal traffic conditions due to accidents or the like occurred. Therefore, the reconstruction error is generally within a constant value. An accident occurred from 15:00 to 16:35 on November 17. The reconstruction error on November 17 is large around 5:00, 6:00, 15:00, and 22:00. However, around 5:00, 6:00, and 22:00, the reconstruction error does not take a large value for a long time, and the reconstruction error fluctuates for a short period of time. These fluctuations in the reconstruction error are noise and do not indicate an abnormal traffic state. In contrast, the reconstruction error continues to take a large value around 15:00. By detecting such fluctuations in the reconstruction error, it is possible to remove the influence of noise and accurately detect an abnormal traffic state. In this embodiment, by determining whether the reconstruction error exceeds a threshold value multiple times in succession, it is detected that the reconstruction error continues to have a large value.

Referring again to FIG. 5, the video signal output unit 115 outputs a video signal indicating the result of the detection of the traffic abnormality state by the state determination unit 114. The video signal is provided to the display device 108, and the result of the detection of the traffic abnormality state is displayed by the display device 108.

7 is a diagram showing an example of a display screen of the display device according to this embodiment. The display device 108 displays a map (hereinafter referred to as a "traffic condition map") that distinguishes between points where a traffic abnormality occurs and points where a traffic abnormality does not occur. In the example shown in FIG. 7, the traffic condition map includes a three-dimensionally displayed map and a cylindrical mark. The mark indicates the traffic condition at the point of the bottom surface of the cylinder. A normal traffic condition is indicated by a green mark (white in the figure), and a traffic abnormality is indicated by a red mark (shaded in the figure). By displaying such a traffic condition map, the user can visually grasp the point where a traffic abnormality occurs. Note that the diagram shown in FIG. 7 is an example of a traffic condition map, and is not limited to such a display format. For example, a mark indicating a traffic condition may be attached to one or more points on a two-dimensional map. The mark may be color-coded (normal is green, abnormal is red, etc.) or changed in shape (normal is a circle, abnormal is a square, etc.) according to the traffic condition. Instead of indicating the traffic condition with a mark, the traffic condition may be expressed with letters such as "normal" and "abnormal".

[5. Learning Function]
Next, a description will be given of a learning function for creating the learning model 106. The learning function is exerted by a training program 107.

A device that executes a training program 107, such as the traffic condition determination device 10, receives a large amount of traffic condition information for a specific time period on a specific day type from the traffic information DB 6 in order to create a learning model 106 for that time period. For example, a user removes abnormal data from the acquired traffic condition information. In other words, data that indicates normal traffic conditions is selected (screened) from the acquired traffic condition information.

The traffic condition information after screening is normalized, and a training dataset is created that contains a large amount of training data in the same data format as the input data above.

The learning model 106 is created by inputting the training data set into the autoencoder. That is, the device that executes the training program 107 performs machine learning using the training data set and constructs (reconstructs) the learning model 106.

Figure 8 is a schematic diagram showing an example of machine learning according to this embodiment. A device that executes a training program 107 provides a training data set to an autoencoder and executes unsupervised machine learning. In unsupervised machine learning, training data is provided as input to the autoencoder, and weights (connection weights) are adjusted so that an output equal to the input is obtained. This type of machine learning is repeated using multiple training data. The connection weights of the autoencoder are set by machine learning, and a learning model 106 is constructed.

6. Traffic Condition Determination Processing
9 is a flowchart showing the procedure of the traffic condition determination process by the traffic condition determination device 10 according to the present embodiment. The traffic condition determination process shown in FIG.

When the traffic condition determination process starts, the processor 101 selects a target point (step S101). The target point is a point to be detected for an abnormal traffic condition, for example, an intersection Ci. Next, the traffic condition determination device 10 receives traffic condition information for the target point from the traffic information DB 6 (step S102).

The processor 101 classifies the received traffic condition information by day type and time period (step S103). The processor 101 normalizes the elements (traffic jam length, traffic volume, and travel time) included in the classified traffic condition information (step S104). This creates input data that includes normalized traffic jam length, traffic volume, and travel time.

The processor 101 inputs the input data to a learning model corresponding to the day type and time period into which the traffic condition information is classified (step S105). The learning model 106 reduces the dimensions of the input model (encodes it) and then restores (decodes) it to generate output data.

The processor 101 calculates the reconstruction error, which is the difference between the input data and the output data (step S106). Furthermore, the processor 101 compares the reconstruction error with a threshold and determines whether the reconstruction error exceeds a predetermined threshold (step S107). Here, the processor 101 counts the number of times that the reconstruction error has consecutively exceeded the predetermined threshold. For example, if the reconstruction error exceeds the predetermined threshold, the processor 101 increments the value of the counter variable by one, and if the reconstruction error is equal to or less than the predetermined threshold, the processor 101 resets the counter.

The processor 101 determines whether the number of consecutive exceedances is equal to or greater than a specified value (step S108). If the number of consecutive exceedances is less than the specified value (NO in step S108), the processor 101 determines that the traffic conditions are normal (step S109). If the number of consecutive exceedances is equal to or greater than a specified value (YES in step S108), the processor 101 determines that the traffic conditions are abnormal (step S110).

The processor 101 creates a traffic condition map based on the traffic condition determination result. Note that "creating a traffic condition map" here means creating data for displaying the traffic condition map as a video in an internal memory. Furthermore, the processor 101 outputs a video signal of the traffic condition map to the display device 108, and causes the traffic condition map to be displayed on the display device 108 (step S111). This completes the traffic condition determination process.

7. Learning Process
10 is a flowchart showing the procedure of the learning process according to the present embodiment. The learning process shown in FIG.

When the learning process starts, the traffic condition determination device 10 receives traffic condition information for a certain day and a certain time period from the traffic information DB 6 (step S201). For example, when creating a learning model 106 for a period from 3:00 p.m. to 4:00 p.m. on a weekday, the traffic condition determination device 10 acquires traffic condition information for the period from 3:00 p.m. to 4:00 p.m. on past weekdays (Monday, Tuesday, Wednesday, Thursday, and Friday).

The processor 101 removes abnormal data from the acquired traffic condition information (step S202). Furthermore, the processor 101 normalizes the elements (traffic jam length, traffic volume, and travel time) included in the traffic condition information from which the abnormal data has been removed (step S203). This creates a training data set that includes a large amount of training data including normalized traffic jam length, traffic volume, and travel time.

The processor 101 provides the training data set to the autoencoder and executes machine learning (step S204). This results in the construction of the learning model 106. This completes the learning process.

8. Modifications
In the above embodiment, the learning model 106 is provided for each day type and each time period, but this is not limited to the above. For example, multiple learning models 106 may be provided according to only the day type, or multiple learning models 106 may be provided according to only the time period. Furthermore, a traffic abnormality may be detected by a single learning model 106 that corresponds to all the day types and time periods.

The traffic condition determination device 10 is not limited to a configuration in which it executes the traffic condition determination process and the learning process. For example, the learning process may be executed by a device different from the traffic condition determination device 10. In this case, the learning model 106 constructed by the device that executed the learning process or a copy thereof may be written to the non-volatile memory of the traffic condition determination device 10 or another device. This makes it possible to provide the learning model 106 to a device that does not have an environment for executing the learning process.

The congestion length, traffic volume, and travel time included in the traffic condition information are configured to be measured by roadside sensors 8, but are not limited to this. For example, a sensor installed in a vehicle may detect the vehicle speed, vehicle position, etc., and probe information including this detection data may be acquired via communication, and the congestion length, traffic volume, travel time, etc. may be calculated based on the acquired probe information.

[9. Effects]
As described above, the traffic condition determination device 10 includes the learning model 106 and the state determination unit 114. The learning model 106 generates output data indicating the traffic condition at a specific point from input data indicating the traffic condition at a specific point. The state determination unit 114 determines whether or not a traffic abnormality occurs based on the input data and the output data. The learning model 106 is an autoencoder including an input layer 106IN, one or more intermediate layers 106MID, and an output layer 106OUT. The autoencoder reduces the dimension of the input data (encodes) and restores (decodes) the input data so as to obtain output data identical to the input data. When input data in a normal traffic condition is given to the learning model 106 trained by various training data, output data in which the input data is restored with high accuracy is generated. The input data reflecting the traffic abnormality when an unexpected event occurs shows a tendency different from the input data in a normal traffic condition. The learning model 106 does not accurately restore the input data reflecting such a traffic abnormality, and generates output data different from the input data. Therefore, traffic abnormalities can be accurately detected based on the input data and the output data.

The input data may include congestion length, traffic volume, and travel time. The input layer 106IN may include a node N11 that accepts congestion length in the input data, a node N12 that accepts traffic volume in the input data, and a node N13 that accepts travel time in the input data. The output data may include congestion length, traffic volume, and travel time. The output layer 106OUT may include a node N31 that outputs congestion length in the output data, a node N32 that outputs traffic volume in the output data, and a node N33 that outputs travel time in the output data. The traffic condition can be accurately indicated by the congestion length, traffic volume, and travel time. Therefore, abnormal traffic conditions can be accurately detected by using such input data.

The traffic condition determination device 10 may include a plurality of learning models 106 corresponding to a plurality of time periods, and a classification unit 112. The classification unit 112 classifies traffic condition information indicating traffic conditions at a specific point according to a plurality of time periods. Input data based on the traffic condition information classified by the classification unit may be input to the learning model 106 corresponding to the same time period as the time period to which the input data corresponds. Traffic conditions tend to differ depending on the time period, such as morning and evening, night, and daytime. Therefore, abnormal traffic conditions can be accurately detected by providing input data classified according to the time period to the learning model 106 created for each time period.

The multiple learning models 106 may correspond to multiple day types, respectively. The classification unit 112 may classify the traffic condition information according to multiple day types. Traffic conditions tend to differ depending on the day type, such as weekdays, Saturdays, Sundays, and holidays. Therefore, abnormal traffic conditions can be accurately detected by providing input data classified according to the day type to a learning model created for each day type.

The traffic condition determination device 10 may further include a normalization unit 113. The normalization unit 113 normalizes traffic condition information indicating the traffic condition at a specific point and generates input data. In this way, by normalizing the traffic condition information, inconsistencies between input data can be suppressed.

The state determination unit 114 may determine whether or not an abnormal traffic state has occurred by comparing the difference between the input data and the output data with a threshold value. In normal traffic conditions, the output data matches the input data well, whereas in abnormal traffic conditions, the output data differs greatly from the input data. Therefore, by comparing the difference between the input data and the output data with a threshold value, an abnormal traffic state can be easily and accurately detected.

The state determination unit 114 may determine that an abnormal traffic condition is occurring when the difference exceeds a predetermined threshold value a specified number of times in succession. When an unexpected incident occurs, an abnormal traffic condition such as congestion continues. Such an abnormal traffic condition may not be resolved in a short period of time. Therefore, by determining whether the difference exceeds a predetermined threshold value a specified number of times in succession, a continuing abnormal traffic condition can be detected.

The traffic condition determination device 10 may further include a display device 108. The display device 108 displays a map that distinguishes between points where an abnormal traffic condition has occurred and points where an abnormal traffic condition has not occurred, based on the determination result of the condition determination unit 114. This allows the user to visually confirm the point where an abnormal traffic condition has occurred.

The learning model 106 may be generated by machine learning to output information that is similar to the training data when the training data indicating a traffic state is given to the input layer. This makes it possible to create a learning model 106 that can be used to detect abnormal traffic conditions.

[10. Supplementary Notes]
The embodiments disclosed herein are illustrative in all respects and are not restrictive. The scope of the present invention is defined by the claims rather than the above-described embodiments, and includes the meaning equivalent to the claims and all modifications within the scope thereof.

REFERENCE SIGNS LIST 1 Traffic signal control system 2 Signal lamp 4 Traffic signal controller 5 Vehicle 6 Traffic information database 8 Roadside sensor 10 Traffic state determination device 101 Processor 102 Non-volatile memory 103 Volatile memory 104 Input/output interface 105 Traffic state determination program 106 Learning model 107 Training program 108 Display device 111 Input unit 112 Classification unit 113 Normalization unit 114 State determination unit 115 Video signal output unit N11 Node (first input node)
N12 node (second input node)
N13 node (third input node)
N2 node N31 node (first output node)
N32 node (second output node)
N33 node (third output node)

Claims (11)

A learning model that generates output data including a congestion length, a traffic volume, and a travel time at a specific point from input data including the congestion length, a traffic volume, and a travel time at the specific point;
a state determination unit that determines whether or not a traffic abnormality state is occurring based on the input data and the output data;
Equipped with
The learning model is
an input layer including a first input node that accepts the congestion length in the input data, a second input node that accepts the traffic volume in the input data, and a third input node that accepts the travel time in the input data;
one or more intermediate layers;
an output layer including a first output node that outputs the congestion length in the output data, a second output node that outputs the traffic volume in the output data, and a third output node that outputs the travel time in the output data;
is an autoencoder that includes
The intermediate layer reduces the dimension of the input data and restores the input data to obtain the output data that matches the input data.
Traffic condition determination device. A plurality of said learning models corresponding to a plurality of time periods, respectively;
A classification unit that classifies traffic condition information indicating a traffic condition at the specific point according to the plurality of time periods;
Equipped with
The input data based on the traffic state information classified by the classification unit is input to the learning model corresponding to the same time period as the time period corresponding to the input data.
The traffic condition determination device according to claim 1. The plurality of learning models correspond to a plurality of day types,
The classification unit classifies the traffic condition information according to the plurality of day types.
The traffic condition determination device according to claim 2. A normalization unit that normalizes traffic condition information indicating a traffic condition at the specific point and generates the input data.
The traffic condition determination device according to any one of claims 1 to 3. the state determination unit determines whether or not the traffic abnormality state has occurred by comparing a difference between the input data and the output data with a predetermined threshold value;
The traffic condition determination device according to any one of claims 1 to 4. The state determination unit determines that the abnormal traffic state is occurring when the difference exceeds the threshold value a specified number of times in succession.
The traffic condition determination device according to claim 5. and a display device that displays a map in which a location where the traffic abnormality condition occurs and a location where the traffic abnormality condition does not occur are distinguished based on a result of the determination by the state determination unit.
The traffic condition determination device according to any one of claims 1 to 6. The learning model is generated by machine learning for outputting information approximating the training data when training data indicating the traffic state is provided to the input layer.
The traffic condition determination device according to any one of claims 1 to 7. generating output data including a congestion length, a traffic volume, and a travel time at a specific point from input data including the congestion length, a traffic volume, and a travel time at the specific point by using a learning model;
determining whether a traffic abnormality occurs based on the input data and the output data;
Including,
The learning model is
an input layer including a first input node that accepts the congestion length in the input data, a second input node that accepts the traffic volume in the input data, and a third input node that accepts the travel time in the input data;
one or more intermediate layers;
an output layer including a first output node that outputs the congestion length in the output data, a second output node that outputs the traffic volume in the output data, and a third output node that outputs the travel time in the output data;
is an autoencoder that includes
The intermediate layer reduces the dimension of the input data and restores the input data to obtain the output data that matches the input data.
A method for determining traffic conditions. A computer program for causing a computer to detect a traffic abnormality, comprising:
The computer includes:
generating output data including a congestion length, a traffic volume, and a travel time at a specific point from input data including the congestion length, a traffic volume, and a travel time at the specific point by using a learning model;
determining whether a traffic abnormality occurs based on the input data and the output data;
Run the command,
The learning model is
an input layer including a first input node that accepts the congestion length in the input data, a second input node that accepts the traffic volume in the input data, and a third input node that accepts the travel time in the input data;
one or more intermediate layers;
an output layer including a first output node that outputs the congestion length in the output data, a second output node that outputs the traffic volume in the output data, and a third output node that outputs the travel time in the output data;
is an autoencoder that includes
The intermediate layer reduces the dimension of the input data and restores the input data to obtain the output data that matches the input data.
Computer program. A learning model for making a computer function , the learning model being constructed by an autoencoder,
an input layer including a first input node that receives a congestion length at a specific point, a second input node that receives a traffic volume at the specific point, and a third input node that receives a travel time at the specific point;
an output layer including a first output node that outputs a congestion length at the specific point, a second output node that outputs a traffic volume at the specific point, and a third output node that outputs a travel time at the specific point;
one or more intermediate layers provided between the input layer and the output layer;
Equipped with
The intermediate layer performs dimensional compression on input data including the congestion length, the traffic volume, and the travel time received by the input layer, and restores output data including the congestion length, the traffic volume, and the travel time output from the output layer so that the output data matches the input data.
A learning model for making computers work .

Images