JP7740313B2 - Data Extraction Device - Google Patents

Description

This disclosure relates to a device that uses a computing device that performs a predetermined calculation based on input data to support efficient calculations and correction of calculation programs when a predetermined calculation is repeated multiple times on a large amount of data.

When developing a computer system (program), the developed system is input with data intended for actual operation, calculations are performed, and the results are checked to see if there are any errors. If an error is found, the program is corrected and the calculations are performed again. This calculation and correction process is repeated until the error is resolved. Once the error is resolved, different data is input and the calculation and correction process is repeated in the same way.

When developing autonomous driving systems, vehicles equipped with autonomous driving systems are tested to determine whether they can be driven safely under various road conditions. If a dangerous situation such as a collision or near miss occurs, the autonomous driving system is modified to avoid the dangerous situation, and then the system is tested again under the same road conditions, and this process is repeated.

In actual road conditions, there are a huge number of possible scenarios, including road environments such as road width, number of lanes, intersection shapes, and the presence or absence of traffic lights, as well as the positions and numbers of nearby vehicles, pedestrians, cyclists, etc., and the behavior of traffic participants such as direction and speed. Autonomous driving systems are required to be able to respond to all road conditions, and for this reason, numerous tests are conducted repeatedly.

Autonomous driving systems are sometimes tested by running a real vehicle on real roads or test courses, but in reality, it is extremely difficult to test a real vehicle under every possible scenario. Therefore, consideration is being given to evaluating safety by virtually running a vehicle equipped with an autonomous driving system on a simulator that recreates various road conditions (Non-Patent Document 1).

When conducting tests on a simulation device, various road conditions, known as scenarios, are prepared. A large number of scenarios are automatically created by changing and combining parameters such as road geometry (intersections, T-junctions, straight roads, presence or absence of traffic lights, etc.), traffic participants (passenger cars, trucks, buses, bicycles, pedestrians, etc.), and the behavior of each traffic participant (speed, trajectory, etc.).

During testing, the autonomous vehicle is virtually driven through each prepared scenario, and safety is evaluated and the autonomous driving program is repeatedly modified.

"FY2018 Research, Development and Demonstration Project for Social Implementation of Advanced Automated Driving Systems: Research and Development Project for Building Safety Evaluation Technology for Automated Driving Systems," Japan Automobile Research Institute, March 2019 (https://www.meti.go.jp/meti_lib/report/H30FY/000351.pdf)

Depending on the system, the amount of data to be verified can be enormous, requiring a huge amount of time and effort to verify the program. Thousands to tens of thousands of scenarios are created for testing autonomous driving systems, even for just a single intersection. Testing all of these thousands to tens of thousands of scenarios and making corrections to the autonomous driving system requires a huge amount of time and effort.

The purpose of this disclosure is to provide a device that can perform processing efficiently and in a short amount of time when repeatedly performing predetermined calculations based on a large amount of data.

The data extraction device disclosed herein includes a data acquisition unit that acquires multiple pieces of scenario data to be input into a test device that runs a simulation of an autonomous driving device based on input scenario data and outputs the results; a grouping criterion acquisition unit that acquires grouping criterion that is a criterion for dividing the multiple pieces of scenario data into groups; a grouping unit that divides the multiple scenario data into one or more groups according to the grouping criterion; and a group representative data extraction unit that extracts group representative data from the groups that is scenario data that represents the groups, wherein the scenario data includes road shape, traffic participants, and the behavior of the traffic participants as parameters, and the grouping criterion is a criterion for grouping together scenario data that include parameters that are predicted to cause the vehicle to behave in the same way when the simulation is run .

According to the present disclosure, by performing calculations on extracted group representative data, calculation results can be confirmed and programs can be modified efficiently without performing calculations on all data. In particular, by grouping or classifying data that is expected to produce substantially identical or similar calculation results and performing calculations on group representative data or classification representative data, calculation results can be utilized more efficiently.

1 shows an example of a functional configuration. Another example of the functional configuration will be shown. The system configuration is shown below. 1 shows the hardware configuration of a data extraction device. A flowchart of grouping is shown. A classification flowchart is shown. Classify, group, and show data relationships. 1 shows an example of a system configuration. The first scenario is shown. A second scenario is shown. A third scenario is shown. Provide a description of the near miss. A fourth scenario is shown. A fifth scenario is shown. A flowchart of grouping is shown. A classification flowchart is shown. An explanation of the collision direction is shown.

The data extraction device according to the embodiment of the present disclosure will be described in detail below with reference to the drawings. Note that each drawing is a schematic diagram and is not necessarily an exact illustration. Furthermore, in each drawing, substantially identical components are designated by the same reference numerals, and duplicate explanations may be omitted or simplified.

Figure 1 shows an example of the functional configuration of a data extraction device disclosed herein. The data acquisition unit acquires data to be input to the calculation device. There are no limitations on the type or quantity of data, but in the case of scenario data for an autonomous driving device, thousands to tens of thousands of sets of scenario data may be acquired.

The grouping criteria acquisition unit acquires criteria for grouping the acquired data. The grouping criteria are not limited, but may be, for example, criteria for grouping data that is predicted to produce the same results when calculations are performed based on the data. If the data is scenario data, scenario data that only differs in parameters unrelated to the vehicle's driving, such as the presence or absence of vehicles driving at a distance from the vehicle, is predicted to produce the same results in simulations based on this data, and therefore such scenario data can be grouped together.

The grouping unit divides the acquired data into groups based on the acquired grouping criteria. The number of groups created depends on the acquired data and grouping criteria, but it may create a group containing just one piece of data, or it may create a group containing hundreds of pieces of data.

The group representative data extraction criteria acquisition unit acquires criteria for extracting representative data from a group. The group representative data extraction criteria are not limited, but may be, for example, criteria such as small data size or data created early.

The group representative data extraction unit extracts group representative data from each group according to the group representative data extraction criteria. There is no limit to the number of data extracted from each group, but in order to efficiently perform calculations by the calculation device, it is preferable to have as few data as possible, and in some cases, at least one data may be sufficient.

There are no limitations on the type or purpose of the computing device into which the data is input, but it could, for example, be a test device that simulates an autonomous driving system based on scenario data. The test device runs a simulation of the scenario data to verify whether the autonomous driving system can drive safely under various road conditions. If a collision or near miss occurs as a result of the simulation, this indicates that there is a problem with the autonomous driving system, and the program for the autonomous driving system is modified and the simulation is run again.

When the amount of data becomes enormous, the number of calculations also becomes enormous. With the data extraction device disclosed herein, data that is expected to produce the same calculation results is grouped together, and calculations are performed on group representative data extracted from each group, making it possible to reduce the number of calculations. It is expected that the calculation results will be different for each extracted group representative data, and in some cases, performing calculations only on this extracted data can be considered to be essentially the same as performing calculations on all data.

Figure 2 shows another example of the functional configuration of the data extraction device of the present disclosure. The example of Figure 2 differs from the example shown in Figure 1 in that similar groups are sorted into the same classification and representative classification data is extracted from each classification. The data acquisition unit, grouping criteria acquisition unit, grouping unit, group representative data extraction criteria acquisition unit, and group representative data extraction unit are the same as those shown in Figure 1, so detailed explanations will be omitted.

The classification criteria acquisition unit acquires classification criteria, which are criteria for classifying groups. The classification criteria are not limited, but may be, for example, criteria that classify groups containing data that are predicted to produce similar results when calculations are performed based on the data in the same category. If the data is scenario data, for example, when the subject vehicle collides with another vehicle, groups of scenario data that differ only in that the other vehicle's position is slightly shifted are predicted to require the same type of correction to the autonomous driving device, and so the groups can be classified in the same category. In other words, if it is predicted that corrections made to accommodate the scenario data included in one group will also be possible to accommodate the scenario data included in the other group, both groups can be classified in the same category.

The classification unit classifies the groups based on the classification criteria obtained. The number of classifications depends on the groups and the classification criteria obtained, but there are cases where a classification containing only one group is created, and cases where a classification containing hundreds of groups is created.

The classification representative data extraction criteria acquisition unit acquires criteria for extracting representative data from a classification. The classification representative data extraction criteria are not limited, but may be, for example, criteria such as small data size or data created early.

The classification representative data extraction unit extracts classification representative data from each classification according to the classification representative data extraction criteria. There is no limit to the number of data extracted from each classification, but in order to efficiently perform calculations by the calculation device, it is preferable to have as few data as possible, and in some cases, at least one data may be sufficient.

As with the example shown in Figure 1, the type and use of the computing device are not limited, but it may be, for example, a test device that simulates an autonomous driving device based on scenario data. In the data extraction device disclosed herein, scenario data that is predicted to produce the same simulation results is grouped together, and groups that are predicted to produce similar simulation results are classified into the same category. Then, a simulation is performed on representative category data extracted from each category, making it possible to reduce the number of simulations.

It is estimated that the results of calculations for each extracted representative data category will be different, and that by performing calculations on this extracted data, program modifications will generally be completed without having to perform calculations on the representative data for all groups.

Once calculations and program modifications have been completed for all classification representative data, calculations are performed on group representative data. Since group representative data for groups in the same classification are expected to produce similar calculation results, if there is no malfunction in the calculation device for one group representative data, there is a high probability that there will be no malfunction for other group representative data.

Once calculations and program modifications have been completed for all group representative data, calculations will finally be performed for all remaining data. Since it is assumed that the remaining data will yield the same results as the group representative data for each group, it is unlikely that program modifications will be required due to calculations for all remaining data.

Figure 3 shows an example of a system in which a predetermined calculation is performed by a calculation device using data extracted by the data extraction device of the present disclosure. Data creation device 3 creates data to be input to calculation device 5. Data extraction device 1 extracts and outputs data to be input to calculation device 5 from the data created by data creation device 3. Calculation device 5 performs a predetermined calculation using the data extracted by data extraction device 1. If a defect is discovered in the calculation program as a result of the calculation, calculation program correction device 7 corrects the program. After the program has been corrected, calculation device 5 performs the predetermined calculation again to check for any defects.

If the calculation program has been corrected and no defects have been found in all of the data extracted by the data extraction device 1, the calculation device 5 repeats the specified calculations on the remaining data created by the data creation device 3. If a defect is found in the program during this process, the calculation program correction device 7 corrects the program.

Figure 4 shows an example of the hardware configuration of the data extraction device 1. Connected to the CPU 51 are an SSD 53, memory 55, DVD-ROM drive 57, display 59, keyboard/mouse 61, and communication circuit 63. The communication circuit 63 is a circuit for connecting to a network such as the Internet or an in-house LAN. The SSD 53 stores an operating system 65 and an extraction program 67. The extraction program 67 performs its functions in cooperation with the operating system 65. The SSD 53 also stores data 69 created by the data creation device 3, group data 71 obtained by grouping the data 69, and classified data 73 obtained by classifying the group data 71.

Next, the procedure for grouping N pieces of data input to a predetermined calculation device 5 will be described with reference to the flowchart in Figure 5. When grouping begins (step S1), the CPU 51 sets arbitrarily selected data as data 1 with i = 1 (step S2) and starts a processing loop (steps S3 and S4). Next, the CPU 51 determines whether there is an existing group that is determined to be the same group as data 1. When i = 1, there is no existing group, so the determination result is No (step S5). Then, data 1 is sorted into group 1 with the group number set to 1 (step S7).

Next, returning to the beginning of the loop (step S8), the CPU 51 sets the next data, i=2, as data 2 and starts the processing loop (steps S3, S4). The CPU 51 determines whether there is an existing group that is determined to be the same group as data 2 (step S5). Data 1 is included in existing group 1. If it is predicted that the same result will be obtained when data 1 and data 2 are input into a predetermined calculation device 5, the answer is Yes, and data 2 is sorted into group 1 (step S6). On the other hand, if it is predicted that the same result will not be obtained when data 1 and data 2 are input into a predetermined calculation device 5, the answer is No, and data 2 is sorted into group 2 with group number 2 (step S7).

Then, the loop is repeated for all data, and the process ends when all data has been sorted into one of the groups (step S9).

Next, the procedure for further classifying the M groups will be described with reference to the flowchart in Figure 6. Once the data has been grouped, classification begins (step S11). First, the CPU 51 sets group 1 as group i = 1 (step S12) and starts a processing loop (steps S13 and S14). Next, the CPU 51 determines whether there is an existing classification that is determined to be the same classification as group 1. When i = 1, there is no existing classification, so the determination result is No (step S15). Then, the classification number is set to 1 and group 1 is sorted into classification 1 (step S17).

Next, the process returns to the beginning of the loop (step S13), and the CPU 51 sets the next group 2 as group i=2 and starts the processing loop (steps S13, S14). Next, the CPU 51 determines whether there is an existing classification that is determined to be the same classification as group 2 (step S15). Group 1 is included in existing classification 1. If it is predicted that similar results will be obtained when the data for group 1 and the data for group 2 are input into a specified calculation device 5, the answer is Yes, and group 2 is sorted into classification 1 (step S16). On the other hand, if it is predicted that identical results will not be obtained when the data for group 1 and the data for group 2 are input into a specified calculation device 5, the answer is No, and group 2 is sorted into classification 2 with a classification number of 2 (step S17).

Then, the loop is repeated for all groups, and the process ends when all groups have been sorted into one of the categories (step S19).

Figure 7 shows a schematic example of the relationship between data, groups, and classifications after all grouping and classification has been completed. Each group contains one or more data items. Each classification contains one or more groups.

Once grouping and classification are complete, the CPU 51 extracts the data to be input to the arithmetic unit 5. In Figure 7, it is assumed that if data 1 to data 3 in group 1 were input into the arithmetic unit 5, the same result would be obtained. Therefore, for example, data 1, which has the smallest data number, is selected from group 1. It is assumed that if data 4 to data 6 in group 2 were input into the arithmetic unit 5, the same result (a different result from data 1 to data 3 in group 1) would be obtained. Therefore, for example, data 4, which has the smallest data number, is selected from group 2. Subsequently, group representative data is extracted from all groups.

Next, the CPU 51 extracts representative data from each classification. When extracting representative data for classification 1, for example, data 1 from group 1, which has the smallest group number, is extracted as representative data for classification 1 from data 1 extracted from group 1, data 4 extracted from group 2, and data 7 extracted from group 3. Similarly, representative data is extracted from the other classifications.

The criteria for extracting group representative data and classification representative data may be, for example, to select data with a small size. This reduces the computational load and allows for efficient work such as program modification.

Below, we will explain a specific example of the data extraction device of the present disclosure, in which the computing device 5 is a test device that simulates an autonomous driving device.

Figure 8 shows an example of a system in which a simulation of an autonomous driving device is performed by a test device using scenario data extracted by the data extraction device disclosed herein. The scenario creation device 13 creates a scenario for testing a vehicle equipped with an autonomous driving device by driving it in a simulation. A large number of scenarios are automatically created by changing and combining parameters such as road shape (intersections, T-junctions, straight roads, presence or absence of traffic lights, etc.), traffic participants (passenger cars, trucks, buses, bicycles, pedestrians, etc.), and the behavior of each traffic participant (speed, trajectory, etc.). At this stage, the scenario is created assuming that the vehicle is not driven autonomously, but rather travels at a predetermined speed and trajectory.

The scenarios created are divided into risk scenarios in which the vehicle experiences a dangerous situation such as a collision or near miss, and safety scenarios in which no dangerous situation occurs. Risk scenarios are scenarios in which the vehicle should avoid collisions or near misses if it were to drive autonomously.

The data extraction device 11 groups and categorizes the risk scenarios according to criteria such as the other party involved in the collision and the direction of the collision. As will be described in detail below, the grouping and categorization are performed by categorizing scenarios that are predicted to result in the same or similar behavior of the vehicle's autonomous driving device. A group representative scenario is then extracted from each group, and a category representative scenario is extracted from each category.

The test device 15 performs tests using the classification representative scenario data extracted by the data extraction device 11. Specifically, the classification representative scenario data is used to automatically drive the vehicle, and a simulation is performed to see whether a collision or near miss with another vehicle can be avoided. If a collision or near miss occurs, it is determined that the automatic driving device needs to be modified.

The automatic driving program correction device 17 corrects the automatic driving device when correction is necessary. It analyzes collisions and near misses that occur and makes the necessary corrections to the automatic driving device's program.

Figure 9 shows a schematic example of a scenario created by the scenario creation device 13. The first scenario shown in Figure 9 depicts a scene in which the host vehicle 100 is traveling on a travel road 41 and is approaching an intersection 43 with a cross road 42 that crosses ahead. Ahead of the host vehicle 100, vehicle 22, which has passed the intersection 43, is traveling in the same direction as the host vehicle on the travel road 41. In the opposite lane of the travel road 41, vehicles 23 and 25 are traveling in the opposite direction from the host vehicle 100. Vehicle 23 turns left at the intersection 43 and changes course onto the cross road 42. Vehicle 25 has passed the intersection 43 and is traveling away from the host vehicle 100.

At the crossroad 42, vehicle 24 is approaching intersection 43. Vehicle 24 approaches intersection 43 from the left side as viewed from vehicle 100 and stops just before intersection 43.

Vehicle 100 enters intersection 43 and continues traveling straight ahead, upward in the figure. In this scenario, vehicle 100 travels safely without colliding with or near-missing other vehicles 22-25.

Figure 10 shows a schematic diagram of another example of a created scenario. In the second scenario shown in Figure 10, the behavior of vehicle 100 and vehicles 22 to 25 is the same as in the first scenario. In this second scenario, vehicle 21 is also traveling across intersecting road 42 from the right to the left as viewed from vehicle 100.

Vehicle 21 entered intersection 43 from the right side of vehicle 100 and attempted to cross directly in front of vehicle 100, but vehicle 100 collided with vehicle 21 slightly to the rear of its left side. Because a collision occurred in this second scenario, the second scenario is considered a "dangerous scenario."

Figure 11 schematically shows another example of a created scenario. In the third scenario shown in Figure 11, the behavior of host vehicle 100 and vehicles 22 to 25 is the same as in the first and second scenarios. However, vehicle 21 enters and passes through intersection 43 earlier than in the second scenario, and therefore does not collide with host vehicle 100, instead passing by just before it continues traveling. If host vehicle 100 and vehicle 21 approach each other to within a predetermined distance, it is determined that a near miss between host vehicle 100 and vehicle 21 has occurred, and the third scenario is designated a "risk scenario."

Here, a near miss refers to a situation where the vehicle 100 and another traffic participant are within a predetermined distance, as shown in Figure 12. Safety areas 101, which are larger than the vehicle body by a predetermined length, are set on the front, rear, left, and right sides of the vehicle 100. The safety area 101 is an area that is deemed dangerous if another traffic participant approaches within this area, and if another traffic participant enters this area, it is determined to be a near miss.

Near misses are also determined by TTC (Time to Collision), which indicates the time until collision if the current behavior is continued. In this disclosure, near misses are determined using two indicators: the safety area and TTC.

Figure 13 schematically shows another example of a created scenario. In the fourth scenario shown in Figure 13, the behavior of host vehicle 100, vehicles 21, 23 to 25 is the same as in the second scenario. On the other hand, vehicle 22, which was traveling in front of host vehicle 100 in the second scenario, is not present in this fourth scenario. In the fourth scenario, the behavior of host vehicle 100 and vehicle 21 is the same as in the second scenario, so host vehicle 100 and vehicle 21 collide, just as in the second scenario. The fourth scenario is then designated as a "dangerous scenario."

Figure 14 schematically shows another example of a created scenario. In the fifth scenario shown in Figure 14, the behavior of host vehicle 100 and vehicles 22 to 25 is the same as in the second scenario. However, vehicle 21 enters intersection 43 later than in the second scenario. Host vehicle 100 then collides with the left side of vehicle 21, slightly forward. In other words, the location of collision for vehicle 21 differs between the second and fifth scenarios. The fifth scenario is therefore designated as a "dangerous scenario."

In the first scenario shown in Figure 9, the host vehicle 100 safely passes through the intersection 43, so there is no need to use the first scenario to simulate whether the host vehicle 100 can be driven automatically to avoid collisions or near misses with other vehicles 21 to 25. On the other hand, for the second to fifth scenarios, which are considered dangerous scenarios, it is necessary to simulate whether the host vehicle 100 can be driven automatically to avoid collisions or near misses with other vehicles 21 to 25.

Next, the procedure for grouping scenario data will be explained with reference to the flowchart in Figure 15, using the second to fifth scenarios, which are considered risk scenarios, as an example. When grouping begins (step S21), the CPU 51 starts a processing loop with the second scenario as the scenario data with i = 1 (steps S22, S23, S24). Next, the CPU 51 determines whether there is an existing group that is determined to be the same group as the second scenario, and since there is no group yet, the determination result is No (step S25). Then, the group number is set to 1 and the second scenario is sorted into group 1 (step S27).

Now, step S25 will be explained. In this step, scenario data is sorted into groups according to predetermined criteria. Scenarios in which the circumstances of a collision or near miss that occurred to the host vehicle are the same and only differ in parts unrelated to the collision or near miss are assumed to result in the same behavior to be taken by the host vehicle's automated driving device, and such scenario data is classified as the same group. Specifically, if the presence or absence of a collision with the host vehicle, the collision partner, the collision location, the collision speed, the minimum TTC, and the minimum distance to the safety area are all the same, they are determined to belong to the same group. Note that the criteria for determining whether scenarios belong to the same group are determined appropriately depending on factors such as the time required for the simulation, and are not limited to the above.

Next, returning to the beginning of the loop (step S28), the CPU 51 sets the third scenario as the scenario data for i=2 and starts the processing loop (steps S23, S24). Next, the CPU 51 determines whether there is any existing group that is determined to be the same group as the third scenario (step S25). The second scenario is included in existing group 1. Comparing the second scenario and the third scenario, in the second scenario, the host vehicle 100 collides with the left side of vehicle 21, slightly to the rear, whereas in the third scenario, the host vehicle 100 has a near miss with vehicle 21 but does not actually collide. Therefore, the same group determination is No (step S25). The group number is then set to 2 and the third scenario is sorted into group 2 (step S27).

Next, returning to the beginning of the loop (step S28), the CPU 51 starts a processing loop with the fourth scenario as the scenario data for i=3 (steps S23, S24). Next, the CPU 51 determines whether any existing groups are determined to be in the same group as the fourth scenario (step S25). The second scenario is included in existing group 1. Comparing the second scenario with the fourth scenario, in both the second and fourth scenarios, the host vehicle 100 collided with the left side of vehicle 21, slightly to the rear. The only difference between the second and fourth scenarios is the presence or absence of vehicle 22 traveling in front of the host vehicle 100, but the presence or absence of vehicle 22 does not affect the collision between the host vehicle 100 and vehicle 21. Therefore, the fourth scenario is determined to be in the same group as the second scenario (Yes) (step S25). The fourth scenario is then sorted into group 1 (step S26).

Next, returning to the beginning of the loop (step S28), the CPU 51 sets the fifth scenario as the scenario data for i=4 and starts the processing loop (steps S23, S24). Next, the CPU 51 determines whether any existing groups are determined to be the same group as the fifth scenario (step S25). Existing group 1 contains the second and fourth scenarios. Comparing the second and fifth scenarios, in both the second and fifth scenarios, the host vehicle 100 collided with the left side of vehicle 21. However, while in the second scenario the host vehicle 100 collided slightly to the rear of the left side of vehicle 21, in the fifth scenario the host vehicle 100 collided slightly to the front of the left side of vehicle 21. Therefore, since the collision locations are different, the fifth scenario is determined to be different from group 1.

Next, the third scenario is included in the existing group 2. Comparing the third scenario with the fifth scenario, in the third scenario, the host vehicle 100 has a near miss with vehicle 21 but does not collide, whereas in the fifth scenario, the host vehicle 100 collides with the left side of vehicle 21, slightly forward. Therefore, the fifth scenario is determined to be different from group 2 as well. Therefore, the same group determination is No (step S25), and the group number is set to 3, sorting the fifth scenario into group 3 (step S27).

As a result of the above, the second and fourth scenarios are sorted into group 1, the third scenario into group 2, and the fifth scenario into group 3.

Once all scenario data has been grouped (step S28), the grouping process ends (step S29).

Next, the procedure for further classifying each group will be explained with reference to the flowchart in Figure 16, using as an example groups 1 to 3, into which scenarios 2 to 5 have been sorted. Once grouping of scenario data has been completed, classification begins (step S31). The CPU 51 starts a processing loop with group 1 as group i = 1 (steps S32, S33, S34). Next, the CPU 51 determines whether the scenario data in group 1 is scenario data in which a collision occurred (step S35). Since the second and fourth scenarios sorted into group 1 are scenarios in which a collision occurred (Yes), the CPU 51 then performs a collision classification determination (step S36). Since there is no existing classification yet, the determination is No (step S36), and group 1 is classified into classification 1, with a classification number of 1 (step S38).

Now, step S36 will be described. In this step, groups are classified according to predetermined criteria. Scenarios in groups with similar collision situations involving the host vehicle are assumed to result in the same or similar behaviors being taken by the host vehicle's automated driving device, and such groups are classified as the same. Specifically, if the trajectory of the collided vehicle, the type of target, and the direction of collision are all the same, they are determined to be in the same category. Note that the criteria for determining whether a group is classified as the same are determined appropriately depending on factors such as the time required for the simulation, and are not limited to the above.

Next, the process returns to the beginning of the loop (step S42), and the CPU 51 starts a processing loop with group 2 as the group i=2 (steps S33, S34). Next, the CPU 51 determines whether the scenario data for group 2 is scenario data in which a collision occurred (step S35). Since the third scenario sorted into group 2 is a near-miss scenario (No), a near-miss classification determination is then performed (step S39). As there is no existing classification for near-misses, the determination is No (step S39), and group 2 is classified into classification 2 with a classification number of 2 (step S41).

Now, step S39 will be described. In this step, groups are classified according to predetermined criteria. It is assumed that scenarios in groups with similar near-miss situations that occurred to the host vehicle will result in the same or similar behavior to be taken by the host vehicle's automated driving device, and such groups are classified as the same. Specifically, if the trajectory of the near-miss target, the type of target, and the type of near-miss are all the same, they are determined to be in the same category. Here, the near-miss type refers to a near-miss from the TTC perspective, a near-miss from the perspective of a safety area, or a near-miss from both the TTC and safety area perspectives. Note that the criteria for determining whether a group is classified as the same are determined appropriately depending on factors such as the time required for the simulation, and are not limited to the above.

Figure 17 shows a schematic explanation of collision direction. The angle forward from the center O of the host vehicle 100 is set to 0 degrees, and the right front corner is at an angle of θ from the center O. If the vehicle collides with a part of the host vehicle 100 between -θ and +θ, it is determined that the collision occurred at the front. If the vehicle collides with the other vehicle between +θ and 180-θ, it is determined that the collision occurred at the right. If the vehicle collides with the other vehicle between -(180-θ) and -θ, it is determined that the collision occurred at the left. If the vehicle collides with the other vehicle between 180-θ or more or -(180-θ) or less from the host vehicle 100, it is determined that the collision occurred at the rear.

Returning to the flowchart, the CPU 51 then returns to the beginning of the loop (step S42), and starts a processing loop with group 3 as the group i=3 (steps S33, S34). Next, the CPU 51 determines whether the scenario data of group 3 is scenario data in which a collision occurred (step S35). Since the fifth scenario sorted into group 3 is a scenario in which a collision occurred (Yes), a collision classification determination is then performed (step S36).

Group 1 is included in the existing classification 1. Comparing the fifth scenario in group 3 with the second scenario in group 1, in the second scenario, host vehicle 100 collided with vehicle 21 slightly to the rear on the left side, whereas in the fifth scenario, host vehicle 100 collided with vehicle 21 slightly to the front on the left side. In both cases, host vehicle 100 collided with the front part of host vehicle 100, so the collision direction is the same. Therefore, group 1 and group 3 are determined to be in the same classification (Yes), and group 3 is classified as classification 1 (step S37).

The collision classification determination will be explained further. A collision classification determination is performed for the i-th group, and if the result is No, a new classification number is assigned and the group is classified (step S38).

The near-miss classification determination will be explained further. A near-miss classification determination is performed for the i-th group (step S39). If the answer is Yes, the existing classification number is assigned and classified (step S40). If the answer is No, a new classification number is assigned and classified (step S41).

Next, we will explain the procedure for extracting group representative scenario data and classification representative scenario data from grouped and categorized scenario data. The group representative scenario data extraction criteria for extracting group representative scenario data can be set arbitrarily, for example, by selecting from the lowest scenario number and extracting the selected scenario with the smallest number of traffic participants. Group representative scenario data is extracted from each group according to the set group representative scenario data extraction criteria.

Next, the criteria for extracting representative scenario data are set separately for the classification of risk scenarios involving collisions and risk scenarios involving near misses. For example, in the case of classification of risk scenarios involving collisions, the criteria can be arbitrarily set by selecting the scenario with the highest vehicle speed at the time of collision, selecting the scenario with the fewest traffic participants from among the selected scenarios, and then extracting the scenario with the smallest group number from among the selected scenarios. For example, in the case of classification of risk scenarios involving near misses, the criteria can be arbitrarily set by selecting the scenario with the smallest near miss distance, selecting the scenario with the fewest traffic participants from among the selected scenarios, and then extracting the scenario with the smallest group number from among the selected scenarios.

The extracted classification representative scenario data is data extracted from groups that are predicted to produce similar simulation results, and it is predicted that by performing a simulation on this extracted scenario data, adjustments to the autonomous driving device will be largely completed without having to perform a simulation on the representative scenario data for all groups.

After running a simulation for all classification representative scenario data, a simulation is run for group representative scenario data. Group representative scenario data for groups in the same classification is expected to produce similar simulation results, so if an autonomous driving device can drive safely using one group representative scenario data, there is a high probability that it will also be able to drive safely using other group representative scenario data.

Once simulations have been completed for all group representative scenario data, simulations will finally be performed for all remaining scenario data. It is assumed that the remaining scenario data will produce results identical to the group representative scenario data for each group, so it is unlikely that simulations for all remaining scenario data will require major program modifications.

As described above, when a huge amount of scenario data is created, the scenario data can be grouped and classified, representative scenario data for each group and representative scenario data for each category can be extracted, simulations can be performed based on the representative scenario data for each category, and then simulations can be performed based on the representative scenario data for each group, thereby efficiently modifying and developing programs for autonomous driving devices.

This disclosure is particularly applicable when the same calculation is repeatedly performed on a large amount of data.

1 Data extraction device 3 Data creation device 5 Arithmetic device 7 Arithmetic program correction device 21 to 25 Vehicle 41 Traveling road 42 Intersection 43 Intersection 51 CPU
100 Vehicle 101 Safety area

Claims (5)

a data acquisition unit that acquires multiple pieces of scenario data to be input to a test device that executes a simulation of an autonomous driving device based on the input scenario data and outputs the results;
a grouping criterion acquisition unit that acquires grouping criteria that are criteria for dividing the plurality of scenario data into groups;
a grouping unit that divides the plurality of scenario data into one or more groups according to the grouping criteria;
a group representative data extraction unit that extracts group representative data, which is scenario data that represents the group, from the group;
and
the scenario data includes, as parameters, road geometry, traffic participants, and behaviors of the traffic participants;
The grouping criterion is a criterion for grouping scenario data including parameters that are estimated to cause the host vehicle to behave in the same manner when the simulation is executed into the same group.
Data extraction device. a group representative data extraction criteria acquisition unit that acquires group representative data extraction criteria that are criteria for extracting the group representative data from the group;
the group representative data extraction unit extracts the group representative data in accordance with the group representative data extraction criteria;
The data extraction device according to claim 1 . a classification criterion acquisition unit that acquires classification criteria that are criteria for classifying the groups;
a classification unit that classifies the group into one or more classes according to the classification criteria;
a classification representative data extraction unit that extracts classification representative data, which is scenario data that represents a classification, from group representative data of each group belonging to the same classification;
The data extraction device according to claim 2 , further comprising: the classification criterion is a criterion for classifying groups of scenario data including parameters that are estimated to cause the host vehicle to behave similarly when the simulation is executed into the same category.
The data extraction device according to claim 3. The method further includes a classification representative data extraction criterion acquisition unit that acquires classification representative data extraction criteria, which are criteria for extracting the classification representative data from the group representative data extracted from a group of the same classification,
the classification representative data extraction unit extracts the classification representative data in accordance with the classification representative data extraction criteria;
5. The data extraction device according to claim 4.