JP2023164047A - Evaluation data preparation device, evaluation data preparation method, and evaluation data preparation program - Google Patents

Abstract

To prepare more evaluation data used for evaluating performance of driving support with a small processing amount.SOLUTION: An evaluation data preparation device for preparing evaluation data for evaluating a driving support function of a vehicle includes: a scenario setting unit for setting a scenario for reproducing travelling of a vehicle; a sensor data preparation unit for processing the scenario set by the scenario setting unit with a sensor model which is modelled after a sensor of the vehicle and preparing sensor data, which is the information detected by the sensor model; and an actual data conversion unit for associating the sensor data prepared by the sensor data preparation unit with actual data, which is the data acquired by the sensor of the vehicle, converting the sensor data on the basis of the associated actual data, and preparing conversion actual data, which is the evaluation data.SELECTED DRAWING: Figure 1

Description

The present disclosure relates to an evaluation data creation device, an evaluation data creation method, and an evaluation data creation program.

Driving support functions such as automatic driving and collision avoidance have been developed to support drivers who drive vehicles. The driving support function determines which support to perform based on data obtained from various sensors installed in the vehicle. The driving support function needs to be compatible with various environments in which the vehicle runs.

Evaluation methods for evaluating the performance of driving support functions have also been proposed. Patent Document 1 describes that the driving environment of a plurality of vehicles is reproduced through simulation, the reproduced information is input to each vehicle, and a driving support function is evaluated based on the executed behavior.

JP 2017-105453 Publication

The device disclosed in Patent Document 1 can perform analysis in various situations by performing analysis through simulation. However, since it is a simulation, there is a deviation from the data that can be obtained when the vehicle is actually driving, so there is a limit to the improvement in reliability when used to evaluate driving support functions.

The device disclosed in Patent Document 1 can perform simulations to simulate various conditions and situations. However, since the evaluation data obtained through the simulation differs from the data obtained when the vehicle is actually driving, it is difficult to increase the reliability of the data to a level suitable for evaluating driving support performance.

On the other hand, by driving an actual vehicle, it is possible to obtain actual driving data instead of simulated data. In this case, in order to obtain reliability suitable for evaluating driving support performance, it is necessary to conduct tests under various conditions and situations. However, there are limits to conducting driving tests using actual vehicles under various conditions and situations. Therefore, it is difficult in reality to obtain data suitable for evaluating driving support performance from actual driving vehicle tests.

In order to solve the above problems, the present disclosure provides an evaluation data creation device, an evaluation data creation method, and an evaluation data creation method that can efficiently create evaluation data with reliability suitable for evaluation of driving support performance. The task is to provide a program.

The present disclosure provides an evaluation data creation device that creates evaluation data for evaluating at least one of automatic driving of a vehicle and a driving support function of the vehicle, the device comprising: a scenario setting unit that sets a scenario for reproducing the driving of the vehicle; , a sensor data creation unit that creates sensor data obtained when the vehicle runs a scenario set by the scenario setting unit, using a sensor model that models the sensor of the vehicle; and the sensor data creation unit The sensor data created in 2. is associated with actual data that is data acquired by the sensor of the vehicle, and the sensor data is converted based on the associated actual data to create converted actual data that is evaluation data. An evaluation data creation device including an actual data conversion section is provided.

The present disclosure also provides an evaluation data creation method for creating evaluation data for evaluating at least one of automatic driving of a vehicle and a driving support function of a vehicle, the method comprising: a step of creating sensor data that is acquired when the vehicle runs the scenario set in the scenario setting section using a sensor model that models the sensor of the vehicle; the step of associating actual data that is data acquired by a sensor of the vehicle, converting the sensor data based on the associated actual data, and creating converted actual data that is evaluation data. Provides a method for creating evaluation data.

The present disclosure also provides an evaluation data creation program that creates evaluation data for evaluating at least one of automatic driving of a vehicle and a driving support function of a vehicle, the program setting a scenario that reproduces the driving of the vehicle. a step of creating sensor data that will be acquired when the vehicle runs the scenario set in the scenario setting section using a sensor model that models the sensors of the vehicle; , a step of associating actual data that is data acquired by a sensor of the vehicle, converting the sensor data based on the associated actual data, and creating converted actual data that is evaluation data. Provides an evaluation data creation program that executes the evaluation data.

With the above configuration, it is possible to create more evaluation data for use in evaluating driving support performance with a smaller amount of processing.

FIG. 1 is a block diagram showing an example of an evaluation data creation device. FIG. 2 is a flowchart showing an example of processing of the evaluation data creation device. FIG. 3 is a flowchart showing an example of the operation of the machine learning section. FIG. 4 is a flowchart showing an example of processing of the evaluation data creation device. FIG. 5 is a flowchart showing an example of processing of the evaluation data creation device. FIG. 6 is a flowchart showing an example of processing of the evaluation data creation device. FIG. 7 is a flowchart showing an example of processing of the evaluation data creation device. FIG. 8 is a block diagram showing an example of an evaluation data creation device. FIG. 9 is a flowchart showing an example of the operation of the machine learning section. FIG. 10 is a flowchart showing an example of processing of the evaluation data creation device.

Embodiments according to the present disclosure will be described in detail below based on the drawings. Note that the present invention is not limited to this embodiment. Furthermore, the constituent elements in the embodiments described below include those that can be easily replaced by those skilled in the art, or those that are substantially the same. Furthermore, the constituent elements described below can be combined as appropriate, and if there are multiple embodiments, it is also possible to combine each embodiment.

<Evaluation data creation device>
FIG. 1 is a block diagram showing an example of an evaluation data creation device. The evaluation data creation device 10 according to the present embodiment creates evaluation data used for evaluation of the driving support function of the ECU (Electronic Control Unit) 6.

The ECU 6 is mounted on the vehicle and executes a driving support function based on information acquired by a sensor mounted on the vehicle. The driving support executed by the ECU 6 is not particularly limited. Examples of driving support include traveling speed control, brake control, steering control, automatic driving control, warning control to the driver, and warning control to other vehicles and surrounding passersby. In other words, at least one of the vehicle's automatic driving function and driving support function is targeted. Sensors installed in the vehicle also contain various information, including information acquired by distance sensors such as LiDAR, temperature sensors, cameras, etc., information acquired from surrounding vehicles, roadside devices, and the cloud via communication, and information input into the vehicle. This includes information on the accelerator, brake, steering, etc., and information on sensors installed in the drive mechanism.

When evaluating the ECU 6, evaluation data created by the evaluation data creation device 10 is input to the ECU 6. When input as sensor data, the ECU 6 processes the input evaluation data and outputs the processing result to the evaluation device 8 as driving support instruction information.

The evaluation device 8 evaluates the driving support function of the ECU 6. The evaluation device 8 is a device that has an arithmetic processing function such as a CPU and a storage function such as a ROM and RAM, and executes an evaluation function using a program. The evaluation device 8 acquires the evaluation data inputted by the evaluation data creation device 10 and the conditions of the evaluation data, processes the evaluation data by the ECU 6, and acquires driving support instruction information outputted. The evaluation device 8 determines whether the timing and content of the driving support executed by the ECU 6 are appropriate for the situation of the scenario of the evaluation data. By evaluating the ECU 6 using the evaluation data, the driving support performance of the ECU 6 can be evaluated. The performance of driving support can be improved by evaluating the ECU 6 and adjusting the functions of the ECU 6 using the evaluation data. Moreover, the evaluation data can also be used as learning data for the ECU 6.

Next, the evaluation data creation device 10 will be explained. The evaluation data creation device 10 reproduces the running state of the vehicle to be evaluated through simulation, and creates evaluation data that simulates data input to the ECU 6 when the vehicle to be evaluated is running.

The evaluation data creation device 10 includes an input section 12, an output section 14, a communication section 15, a calculation section 16, and a storage section 18. The input unit 12 includes an input device such as a keyboard and mouse, a touch panel, or a microphone that collects speech from an operator, and outputs a signal corresponding to an operation performed by the operator on the input device to the calculation unit 16. The output unit 14 includes a display device such as a display, and displays a screen containing various information such as processing results and images to be processed based on the display signal output from the calculation unit 16. Further, the output unit 14 may include a recording device that outputs data on a recording medium. The communication unit 15 uses a communication interface to transmit data. The communication unit 15 communicates with the ECU 6 and the evaluation device 8, and sends and receives data. The communication unit 15 transmits various data and programs acquired through communication with external devices to the storage unit 16 and stores them therein. The communication unit may be connected to an external device via a wired communication line or may be connected to an external device via a wireless communication line.

The calculation unit 16 includes an integrated circuit (processor) such as a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit), and a memory serving as a work area, and can execute various programs using these hardware resources. Execute various processing by. Specifically, the arithmetic unit 16 reads a program stored in the storage unit 18, loads it in the memory, and causes the processor to execute instructions included in the program loaded in the memory, thereby executing various processes. The calculation unit 16 includes a scenario setting unit 22, a sensor data creation unit 24, an actual data conversion unit 26, a correlation processing unit 28, and a machine learning unit 30. Before explaining each part of the calculation section 16, the storage section 18 will be explained.

The storage unit 18 is composed of a nonvolatile storage device such as a magnetic storage device or a semiconductor storage device, and stores various programs and data. The storage unit 18 stores a data creation program 32, a learning program 34, a scenario model 36, an environmental object model 37, a weather model 38, a light source model 40, a vehicle model 42, a sensor model 44, and a correspondence table. 46, actual data 48, and a trained model 50.

The data stored in the storage unit 18 includes a scenario model 36, an environmental object model 37, a weather model 38, a light source model 40, a vehicle model 42, a sensor model 44, a correspondence table 46, and actual data. 48 and a trained model 50. The scenario model 36, the environmental object model 37, the weather model 38, the light source model 40, the vehicle model 42, the sensor model 44, the correspondence table 46, the actual data 48, and the learned model 50 are each There are multiple models available, allowing you to select the model, data, and table you want to use.

The scenario model 36 is a model that sets conditions for simulating vehicle travel. The scenario model 36 is a model that is a combination of models selected from the environmental object model 37, the weather model 38, the light source model 40, and the vehicle model 42. The scenario model 36 can be a model for a certain instant in which no time elapses in order to evaluate the processing of the ECU 6 at a certain moment, or a model in which there is a lapse of time in order to evaluate the processing of the ECU 6 while driving in a predetermined section. It can also be used as a model.

The environmental object model 37 is a model that includes information on the shape of the surroundings of the vehicle. The environmental object model 37 includes information on the shape of the road on which the vehicle travels (straight line, curve, intersection, presence or absence of guardrails, presence or absence of traffic lights), information on surrounding vehicles, information on surrounding passersby, and the like. In other words, the environmental object model 37 includes a model of the three-dimensional shape of the surroundings of the vehicle.

The weather model 38 is a model that reproduces sunny, cloudy, rain, storm, snow, sleet, and vehicle weather. The weather model 38 is a model in which the reproduced situation changes depending on the amount of clouds, rainfall, snowfall, temperature, wind speed, and the like.

The light source model 40 is a model of an object that outputs light around the vehicle, such as the sun, a lighting tower, a building, or an oncoming vehicle.

The vehicle model 42 is a model that reproduces the vehicle to be analyzed and surrounding vehicles. The vehicle model 42 includes information such as the shape of the vehicle, driving performance, and installed sensors.

The sensor model 44 is a model of a sensor installed in a vehicle to be analyzed. The sensor model 44 is a model in which what kind of information is output is set according to surrounding information.

The association table 46 is a model that includes information on the association between sensor data, which is data created through simulation, and actual data obtained by actually driving the vehicle in advance. The correspondence table 46 has information on conditions for performing correspondence. The association table 46 includes a table of conditions for associating actual data when the actual data is associated based on distribution conditions preset based on scenario conditions and acquired sensor data. When using the learned model 50 for association, processing conditions for inputting sensor data to the learned model 50, etc. are associated.

Actual data 48 is data obtained in advance by actually driving the vehicle. Actual data 48 is data detected by a sensor mounted on the vehicle. The actual data 48 also includes information on actual driving of the vehicle, information on environmental objects, information on weather, information on light sources, and information on the vehicle.

The trained model 50 is a model created by the machine learning unit 30 through machine learning. When sensor data is input, the learned model 50 outputs actual data corresponding to the sensor data.

The programs stored in the storage unit 18 include a data creation program 32 and a learning program 34.

The data creation program 32 is a program for creating evaluation data (evaluation data creation program). The data creation program 32 is a program that implements the functions of the scenario setting section 22, the sensor data creation section 24, the actual data conversion section 26, and the association processing section 28.

The learning program 34 is a program that uses machine learning to create a learned model used for part of the processing of the data creation program 32. The learning program 34 is a program that implements the functions of the machine learning section 30. The learning program 34 performs deep learning processing to create a trained model 50 by using the sensor data created by simulation based on the scenario as input and the teacher data with actual data corresponding to the sensor data as output. . As a deep learning model, a GAN (Generative Adversarial Network) or the like can be used. Note that the learning model and machine learning method are not particularly limited, as long as they can convert sensor data into actual data.

The storage unit 18 may install the data creation program 32 and the learning program 34 by reading the data creation program 32 and the learning program 34 recorded on a recording medium, or may install the data creation program 32 and the learning program 34 by providing them on the network. The data creation program 32 and the learning program 34 may be installed by reading the data creation program 32 and the learning program 34.

The functions of each part of the calculation section 16 will be explained. Each part of the calculation unit 16 can be executed by executing a program stored in the storage unit 18. The scenario setting unit 22 acquires the simulation conditions for executing the analysis, and selects models based on the conditions from the environmental object model 37, the weather model 38, the light source model 40, and the vehicle model 42. , create a scenario model that is a simulation model. The simulation conditions for executing the analysis are set based on the user's input results. If a scenario model 36 corresponding to the simulation conditions for performing the analysis is stored in the storage unit 18, the scenario setting unit 22 reads the stored scenario model 36.

The sensor data creation unit 24 creates sensor data when the vehicle runs in the scenario set by the scenario setting unit 22, using a sensor model that models the sensors of the vehicle. The sensor data creation unit 24 executes a simulation based on the scenario model created by the scenario setting unit 22, and creates sensor data that is data detected by a sensor when the vehicle travels based on the scenario model.

The actual data conversion unit 26 associates the sensor data created by the sensor data creation unit 24 with the association processing unit 28, and converts the sensor data into actual data using the results of the association.

The association processing section 28 associates actual data with the sensor data inputted from the actual data conversion section 26 . The association processing unit 28 uses the association table 46 to identify actual data that corresponds to the sensor data. Further, when using the trained model 50, the association processing unit 28 inputs sensor data to the machine learning unit 30 and acquires the output actual data.

The machine learning unit 30 performs deep learning processing to create a trained model 50 by using the sensor data created by simulation based on the scenario as input and the teacher data with real data corresponding to the sensor data as output. do. The machine learning unit 30 also executes processing using the learned model 50. Specifically, the machine learning unit 30 inputs the input sensor data to the learned model 50 and outputs the corresponding actual data.

<Method>
Next, a method for creating evaluation data will be described using FIG. 2. FIG. 2 is a flowchart showing an example of processing of the evaluation data creation device. The process shown in FIG. 2 is a process that does not use machine learning in the association process.

The calculation unit 16 uses the scenario setting unit 22 to determine the scenario of the simulation to be executed (step S12). The scenario setting unit 22 determines a scenario to be executed based on conditions input by the user, conditions set in advance, and the like.

In the calculation unit 16, the scenario setting unit 22 determines the environmental object model, weather model, light source model, and vehicle model to be used based on the scenario, and creates a scenario model (step S14).

The calculation unit 16 inputs the determined scenario model into the sensor model and creates sensor data (step S16). The calculation unit 16 uses the sensor data creation unit 24 to create a sensor model based on the vehicle model 42 . The sensor data creation unit 24 inputs the scenario model into the sensor model, executes a simulation, and creates sensor data detected by the sensor model, that is, data detected by the sensor when the scenario model is used.

The calculation unit 16 determines actual data corresponding to the sensor data (step S18). In the calculation unit 16, the actual data converting unit 26 and the association processing unit 28 process the sensor data using the association table 46 to determine actual data to be associated with the sensor data. In this embodiment, based on the sensor data and the scenario model, actual data with the closest conditions are selected from the actual data associated with the association table 46.

The calculation unit 16 performs a conversion process to actual data (step S20). In the calculation unit 16, the actual data conversion unit 26 converts the sensor data into actual data based on the association determined by the association processing unit 28.

The calculation unit 16 stores the created actual data (step S22). The actual data becomes the evaluation data.

In this embodiment, as described above, the evaluation data can be made into real data by specifying real data for sensor data created in simulation and converting the sensor data into real data.

Here, in the scenario model, for example, the shape of an object is formed by polygons, the characteristics of the object (reflectance, transmittance, texture, etc.) are set as materials, and these are combined to form a three-dimensional model. When setting the scenario model as a simple model, for example, when creating only the information set in the initial settings in the model creation software, this embodiment converts the sensor data into actual data and supplies it to the ECU 6. The evaluation data can be actual data.

Thereby, it is possible to create evaluation data that is close to actual driving without creating a sophisticated model for outputting values close to actual data from the sensor model. Further, since it is possible to create evaluation data that allows highly accurate evaluation while using a simple model as a scenario model, the time required to create evaluation data can be reduced and processing can be simplified. Furthermore, by associating the actual data, it is possible to suppress the deviation of the evaluation data from the actual sensor detection.

In the embodiment described above, the association processing unit 28 determines one piece of real data 48 corresponding to the sensor data, and uses the determined real data as evaluation table data, but the present invention is not limited to this. The evaluation data creation device 10 may create a learned model in a machine learning section, and convert sensor data to actual data using the learned model.

<How to create a trained model>
FIG. 3 is a flowchart showing an example of the operation of the machine learning section. The processing shown in FIG. 3 is executed by the machine learning section. The machine learning unit 30 acquires actual data (step S32). The machine learning unit 30 may acquire actual data stored in the storage unit 18 or may acquire actual data via the communication unit 15.

The machine learning unit 30 creates sensor data (step S34). The machine learning unit 30 executes processing in the scenario setting unit 22 and sensor data creation unit 24 to create sensor data corresponding to actual data. The machine learning unit 30, for example, acquires actual driving conditions of the data and creates sensor data using the created conditions.

The machine learning unit 30 associates the sensor data with the actual data (step S36). That is, the machine learning unit 30 associates the actual data acquired in step S32 with the sensor data created in step S34.

The machine learning unit 30 determines whether the creation of learning data is completed (step S38). For example, the machine learning unit 30 uses as a determination criterion whether there are a set number or more of combinations of sensor data and actual data.

When the machine learning unit 30 determines that the creation of the learning data is not completed (No in step S38), the process returns to step S32 and further creates a combination of sensor data and actual data.

If the machine learning unit 30 determines that the creation of learning data is completed (Yes in step S38), the machine learning unit 30 executes learning of the learning model by using the sensor data as input and the corresponding actual data as output (step S40). ). For example, the machine learning unit 30 uses a part of the combination of sensor data and actual data, which is learning data, as learning data and the rest as verification data, performs learning with the learning data, and verifies with the verification data. .

Through the above processing, the machine learning unit 30 creates a learned model that outputs the corresponding actual data when sensor data is input. Note that the method for creating a trained model is one example and is not limited to this. Although the above embodiment has been described as supervised learning, it is also possible to prepare a learning model and real data and perform unsupervised learning to identify the real data with respect to the sensor data.

<Other examples of how to create evaluation data>
FIG. 4 is a flowchart showing an example of processing of the evaluation data creation device. Among the processes shown in FIG. 4, processes that are the same as those shown in FIG. 2 are given the same reference numerals, and detailed explanations are omitted.

The calculation unit 16 uses the scenario setting unit 22 to determine the scenario of the simulation to be executed (step S12). In the calculation unit 16, the scenario setting unit 22 determines the environmental object model, weather model, light source model, and vehicle model to be used based on the scenario, and creates a scenario model (step S14). The calculation unit 16 inputs the determined scenario model into the sensor model and creates sensor data (step S16).

The calculation unit 16 inputs the sensor data to the trained model and converts it into actual data (step S52).

The calculation unit 16 stores the created actual data (step S22). Actual data becomes evaluation data.

As shown in FIG. 4, the evaluation data creation device 10 converts sensor data into actual data using a trained model created by machine learning (deep learning), without setting various conditions. , data conversion can be performed.

Here, the evaluation data creation device 10 may use a plurality of trained models. When using multiple trained models, the classification conditions are set and learning data is created in which sensor data that satisfies each classification condition is associated with actual data. The machine learning unit 30 creates a learned model for each condition to be classified by executing machine learning for each learning data.

FIG. 5 is a flowchart showing an example of processing of the evaluation data creation device. Among the processes shown in FIG. 5, the same processes as those shown in FIG. 4 are given the same reference numerals, and detailed description thereof will be omitted. FIG. 5 shows a case where trained models are created for each weather and light source condition.

The calculation unit 16 uses the scenario setting unit 22 to determine the scenario of the simulation to be executed (step S12). In the calculation unit 16, the scenario setting unit 22 determines the environmental object model, weather model, light source model, and vehicle model to be used based on the scenario, and creates a scenario model (step S14). The calculation unit 16 inputs the determined scenario model into the sensor model and creates sensor data (step S16).

The calculation unit 16 uses the actual data conversion unit 26 to determine a trained model to be used based on the weather model and the light source model (step S54). The actual data conversion unit 26 acquires the weather and light source conditions of the sensor data based on the information of the scenario model, and determines a learned model corresponding to the acquired conditions. The calculation unit 16 inputs the sensor data to the trained model and converts it into actual data (step S56).

The calculation unit 16 stores the created actual data (step S22). Actual data becomes evaluation data.

In this way, by providing learned data for each condition, it is possible to increase the accuracy of identifying actual data. It is possible to associate actual data that is close to the actual data obtained when actually driving under the conditions of the scenario model.

By using weather and light sources as classification conditions as in this embodiment, it is possible to identify conditions that are difficult to determine using sensor data created from a scenario model. For example, in the case of images, it is possible to prevent the learning results from being mixed depending on whether it is raining or snowing, whether it is cloudy and dark, whether it is dark in the evening, whether it is dark indoors, etc. This makes it possible to further improve the accuracy of conversion of actual data based on sensor data.

The evaluation data creation device may determine learned data to be used based on user input. FIG. 6 is a flowchart showing an example of processing of the evaluation data creation device. Among the processes shown in FIG. 6, the same processes as those shown in FIG.

The calculation unit 16 uses the scenario setting unit 22 to determine the scenario of the simulation to be executed (step S12). In the calculation unit 16, the scenario setting unit 22 determines the environmental object model, weather model, light source model, and vehicle model to be used based on the scenario, and creates a scenario model (step S14). The calculation unit 16 inputs the determined scenario model into the sensor model and creates sensor data (step S16).

The calculation unit 16 uses the actual data conversion unit 26 to determine a trained model to be used based on the input (step S62). The actual data conversion unit 26 determines a corresponding trained model based on the weather and light source information input to the input unit 12. The calculation unit 16 inputs the sensor data to the trained model and converts it into actual data (step S56).

The calculation unit 16 stores the created actual data (step S22). Actual data becomes evaluation data.

The evaluation data creation device 10 can thereby convert the sensor data into actual data of the weather and light source conditions required by the user. In this embodiment, the information is input by the user, but the information may be input from another database.

The evaluation data creation device 10 may set the weather and light source of the scenario determined in step S12 as one standard setting. This makes it possible to reduce the burden of model creation. In addition, since the process selects a trained model based on the user's arbitrary settings, the actual weather and light source data used as evaluation data are converted into actual data of the weather and light source conditions required by the user. be able to.

The evaluation data creation device may classify the learned models for each condition of the environmental object model. FIG. 7 is a flowchart showing an example of processing of the evaluation data creation device. Among the processes shown in FIG. 7, processes that are the same as those shown in FIG.

The calculation unit 16 uses the scenario setting unit 22 to determine the scenario of the simulation to be executed (step S12). In the calculation unit 16, the scenario setting unit 22 determines the environmental object model, weather model, light source model, and vehicle model to be used based on the scenario, and creates a scenario model (step S14). The calculation unit 16 inputs the determined scenario model into the sensor model and creates sensor data (step S16).

The calculation unit 16 uses the actual data conversion unit 26 to determine a learned model to be used based on the environmental object model (step S64). The actual data conversion unit 26 acquires the weather and light source conditions of the sensor data based on the information of the scenario model, and determines a learned model corresponding to the acquired conditions. The calculation unit 16 inputs the sensor data to the trained model and converts it into actual data (step S56).

The calculation unit 16 stores the created actual data (step S22). Actual data becomes evaluation data.

The evaluation data creation device 10 can improve the accuracy of the learned model by classifying the environmental object models by condition and performing learning under similar conditions for people, buildings, trees, etc. Further, it is possible to suppress calculation divergence during learning, and reduce model creation load.

Furthermore, it is preferable that the evaluation data creation device 10 also input an environmental object model as input information during learning. As a result, even if the detection state changes depending on the weather or the light source (day or night), the conditions of the environmental physical model can be processed and associated. Thereby, detection accuracy can be made higher.

<Other examples of evaluation data creation device>
FIG. 8 is a block diagram showing an example of an evaluation data creation device. In addition to the configuration of the evaluation data creation device 10, the evaluation data creation device 10a shown in FIG. 8 includes a weather light source conversion unit 29 and a second learned model 52. Further, the trained model of the evaluation data creation device 10 becomes the first trained model 50. The evaluation data creation device 10 executes the data conversion process twice.

The evaluation data creation device 10a has two trained models, a first trained model 50 and a second trained model 52. The first trained model 50 is a trained model used in the process of converting sensor data into actual data as described above. The first trained model 50 is created by performing learning using a sensor model and actual data using weather and light source conditions as standard conditions.

The second trained model 52 is a trained model used in a process of converting real data into real data with different weather and light source conditions. A method for creating the second trained model 52 will be described later.

The weather light source converter 29 uses the second trained model 52 to create second real data by changing the weather and light source conditions of the real data (first real data) created by the real data converter 26 . The weather light source conversion section 29 changes the weather and light source conditions based on the conditions input through the input section 12 .

FIG. 9 is a flowchart showing an example of the operation of the machine learning section. This is executed by the machine learning section using FIG. The machine learning unit 30 acquires actual data with standard light source and weather conditions (step S72). The machine learning unit 30 may acquire actual data stored in the storage unit 18 or may acquire actual data via the communication unit 15.

The machine learning unit 30 acquires actual data in which the light source and weather are different from standard conditions (step S74). The machine learning unit 30 may acquire actual data stored in the storage unit 18 or may acquire actual data via the communication unit 15.

The machine learning unit 30 associates actual data with the same conditions except for the light source and weather (step S76). That is, the machine learning unit 30 associates the actual data acquired in step S72 with the sensor data created in step S74.

The machine learning unit 30 determines whether the creation of learning data is completed (step S78). For example, the machine learning unit 30 uses whether there are more than a set number of combinations of actual data as a criterion.

When the machine learning unit 30 determines that the creation of the learning data is not completed (No in step S78), the process returns to step S72 and further creates a combination of sensor data and actual data.

When the machine learning unit 30 determines that the creation of the learning data is completed (Yes in step S78), the machine learning unit 30 uses the actual data under the standard conditions as input and the corresponding actual data as output, and starts learning the second learning model. Execute (step S80).

Through the above processing, the machine learning unit 30 creates a second learned model that converts actual data under standard conditions into actual data under different weather and light source conditions.

FIG. 10 is a flowchart showing an example of processing of the evaluation data creation device. The calculation unit 16 uses the scenario setting unit 22 to determine the scenario of the simulation to be executed (step S12). In the calculation unit 16, the scenario setting unit 22 determines an environmental object model and a vehicle model to be used based on the scenario, and creates a scenario model (step S90). The weather model and light source model are based on standard conditions. The calculation unit 16 inputs the determined scenario model into the sensor model and creates sensor data (step S16).

The calculation unit 16 determines the first learned model (step S92). The actual data conversion unit 26 determines a learned model corresponding to the acquired conditions based on the information and input of the scenario model. Note that the first trained model may be one trained model. The calculation unit 16 inputs the sensor data to the first trained model and converts it into first actual data (step S94).

The calculation unit 16 uses the weather light source conversion unit 29 to determine the second trained model to be used based on the input (step S96). The weather light source conversion unit 29 determines a corresponding second learned model based on the weather and light source information input to the input unit 12. The calculation unit 16 inputs the first real data to the second learned model and converts it into second real data (step S98).

The calculation unit 16 stores the created second actual data (step S99). The second actual data becomes evaluation data.

By performing the conversion process twice, the evaluation data creation device 10a can use weather and light source data as reference conditions among the sensor data scenarios created by the sensor data creation unit 24. Thereby, the load of creating sensor data can be reduced. Moreover, the load of creating the first trained model can also be reduced.

The evaluation data creation device 10a changes the weather and light source conditions in the weather light source converter 29, thereby converting the first actual data converted from the sensor data into second actual data with desired weather and light source conditions. can be converted.

Furthermore, the evaluation data creation device 10a may perform machine learning that selects corresponding real data as the second trained model, but based on the teacher data, the first real data is replaced with the second real data. It is preferable to use machine learning to learn the process of converting into . The evaluation data creation device 10a can convert and correct the influence of the weather and light source with higher accuracy by using the second learned model 52 that can convert from the reference state to the desired weather and light source. As a result, various types of second actual data can be created by changing the weather and light source, and a larger amount of evaluation data can be created.

Furthermore, in the embodiment described above, the process of changing the weather and the light source is performed on the actual data, but the process of changing is not limited to the weather and the light source. The actual data may be changed, that is, the actual data may be corrected based on various conditions set in the scenario. This allows the actual data to be closer to the scenario. Furthermore, by being able to perform correction processing, it is possible to create more evaluation data with different conditions.

This disclosure discloses the following inventions. Note that the examples are not limited to the following.
(1) An evaluation data creation device that creates evaluation data for evaluating at least one of automatic driving of a vehicle and a driving support function of the vehicle, the scenario setting unit that sets a scenario that reproduces the driving of the vehicle; a sensor data creation unit that creates sensor data obtained when the vehicle runs a scenario set by the scenario setting unit, using a sensor model that models the sensor of the vehicle; An implementation for associating the created sensor data with actual data that is data acquired by a sensor of the vehicle, converting the sensor data based on the associated actual data, and creating converted actual data that is evaluation data. An evaluation data creation device including a data conversion section.

Thereby, it is possible to create evaluation data that is close to actual driving without creating a sophisticated model for outputting values close to actual data from the sensor model. Further, since it is possible to create evaluation data that allows highly accurate evaluation while using a simple model as a scenario model, the time required to create evaluation data can be reduced and processing can be simplified.

(2) having a storage unit for storing the actual data, comparing the sensor data acquired by the actual data conversion unit with the actual data stored in the storage unit, and determining actual data to be associated with the sensor data; The evaluation data creation device according to (1), including a correlation processing section. By correlating the actual data, it is possible to prevent the evaluation data from deviating from the actual sensor detection.

(3) using learning data in which the sensor data and the actual data are associated, processing the sensor data with a trained model that has been trained with the sensor data as input and the corresponding actual data as output; The evaluation data creation device according to (1), further comprising an association processing unit that determines actual data to be associated with the sensor data. By using a trained model, the load of the mapping process can be reduced.

(4) The association processing unit has a plurality of trained models classified according to scenario conditions, and determines the trained model to be used based on the scenario set by the scenario setting unit. Evaluation data creation device. This makes it possible to increase the accuracy of the trained model. Further, it is possible to suppress calculation divergence during learning, and reduce model creation load.

(5) The evaluation data creation device according to (4), wherein the association processing unit determines a trained model to be used based on at least one of the weather of the scenario and the surrounding light source. By using weather and light sources as classification conditions, it is possible to identify conditions that are difficult to judge using sensor data created from a scenario model. For example, in the case of images, it is possible to prevent the learning results from being mixed depending on whether it is raining or snowing, whether it is cloudy and dark, whether it is dark in the evening, whether it is dark indoors, etc. This makes it possible to further improve the accuracy of conversion of actual data based on sensor data.

(6) The evaluation according to (4) or (5), further comprising an input unit that detects an input, wherein the association processing unit determines a trained model to be used based on the input detected by the input unit. data creation device. Thereby, it is possible to convert the data into actual data that corresponds to the conditions required by the user.

(7) The evaluation data creation device according to any one of (4) to (6), wherein the association processing unit determines a trained model to be used based on the surrounding environment of the vehicle. This makes it possible to increase the accuracy of the trained model. Further, it is possible to suppress calculation divergence during learning, and reduce model creation load.

(8) A weather/light source conversion unit that performs a process of converting at least one of weather and surrounding light sources from the scenario to the converted actual data created by the actual data conversion unit, and creates second converted actual data. The evaluation data creation device according to any one of (1) to (7). Thereby, the load of creating sensor data can be reduced, and more evaluation data can be created.

(9) The weather/light source conversion unit is associated with actual data in which at least one of the weather and the surrounding light sources is different, receives the actual data of the weather and the surrounding light sources of the scenario created by the scenario creation unit, and receives the weather and the surrounding light sources as input. The evaluation data creation device according to (8), which converts the converted actual data into the second converted actual data using a learned model that is trained by outputting actual data that is associated with a different at least one of the surrounding light sources. This allows the actual data to be closer to the scenario. Furthermore, by being able to perform correction processing, it is possible to create more evaluation data with different conditions.

(10) An evaluation data creation method for creating evaluation data for evaluating at least one of a vehicle's automatic driving and a vehicle's driving support function, the step comprising: setting a scenario for reproducing the driving of the vehicle; , using a sensor model that models the sensor of the vehicle, creating sensor data that will be acquired when the vehicle runs the scenario set by the scenario setting section; Creating evaluation data, including the step of associating actual data that is data acquired by a sensor, and converting the sensor data based on the associated actual data to create converted actual data that is evaluation data. Method.

Thereby, it is possible to create evaluation data that is close to actual driving without creating a sophisticated model for outputting values close to actual data from the sensor model. Further, since it is possible to create evaluation data that allows highly accurate evaluation while using a simple model as a scenario model, the time required to create evaluation data can be reduced and processing can be simplified.

(11) An evaluation data creation program for creating evaluation data for evaluating at least one of a vehicle's automatic driving and a vehicle's driving support function, the step of setting a scenario for reproducing the driving of the vehicle; , using a sensor model that models the sensors of the vehicle to create sensor data that will be acquired when the vehicle travels through a set scenario; An evaluation data creation program that executes a process including the step of associating actual data that is data, converting the sensor data based on the associated actual data, and creating converted actual data that is evaluation data. .

Thereby, it is possible to create evaluation data that is close to actual driving without creating a sophisticated model for outputting values close to actual data from the sensor model. Further, since it is possible to create evaluation data that allows highly accurate evaluation while using a simple model as a scenario model, the time required to create evaluation data can be reduced and processing can be simplified.

6 ECU
8 Evaluation device 10 Evaluation data creation device 12 Input section 14 Output section 15 Communication section 16 Arithmetic section 18 Storage section 22 Scenario setting section 24 Sensor data creation section 26 Actual data conversion section 28 Correspondence processing section 30 Machine learning section 32 Data creation section Program 34 Learning program 36 Scenario model 37 Environmental object model 38 Weather model 40 Light source model 42 Vehicle model 44 Sensor model 46 Correspondence table 48 Actual data 50 Learned model

Claims (11)

An evaluation data creation device that creates evaluation data for evaluating at least one of automatic driving of a vehicle and a driving support function of a vehicle, the device comprising:
a scenario setting unit that sets a scenario for reproducing the driving of the vehicle;
a sensor data creation unit that creates sensor data that is acquired when the vehicle runs a scenario set by the scenario setting unit, using a sensor model that models sensors of the vehicle;
The sensor data created by the sensor data creation unit is associated with actual data, which is data acquired by the sensor of the vehicle, and the sensor data is converted based on the associated actual data, thereby converting the sensor data to be evaluation data. An evaluation data creation device including an actual data conversion unit that creates actual data. comprising a storage unit that stores the actual data;
The evaluation data according to claim 1, further comprising a correlation processing unit that compares the sensor data acquired by the actual data conversion unit and the actual data stored in the storage unit, and determines actual data to be associated with the sensor data. Creation device. Using learning data in which the sensor data and the actual data are associated, the sensor data is processed by a trained model that has been trained with the sensor data as input and the corresponding actual data as output, and the sensor data is The evaluation data creation device according to claim 1, further comprising a mapping processing unit that determines actual data to be mapped to. The evaluation data according to claim 3, wherein the association processing unit has a plurality of trained models classified according to scenario conditions, and determines the trained model to be used based on the scenario set by the scenario setting unit. Creation device. The evaluation data creation device according to claim 4, wherein the association processing unit determines the learned model to be used based on at least one of the weather of the scenario and the surrounding light source. has an input section that detects input;
The evaluation data creation device according to claim 4, wherein the association processing unit determines a trained model to be used based on the input detected by the input unit. The evaluation data creation device according to claim 4, wherein the association processing unit determines the trained model to be used based on the surrounding environment of the vehicle. A weather/light source conversion unit that performs a process of converting at least one of weather and surrounding light sources from the scenario to the converted actual data created by the actual data conversion unit to create second converted actual data. 8. The evaluation data creation device according to any one of 1 to 7. The weather/light source conversion unit is associated with actual data in which at least one of the weather and the surrounding light sources is different, and inputs the actual data of the weather and the surrounding light sources of the scenario created by the scenario creation unit, and converts the weather and the surrounding light sources. 9. The evaluation data creation device according to claim 8, wherein the converted actual data is converted to the second converted actual data using a learned model that is trained by outputting associated actual data in which at least one of the following is different. An evaluation data creation method for creating evaluation data for evaluating at least one of a vehicle's automatic driving and a vehicle's driving support function, the method comprising:
setting a scenario that reproduces the driving of the vehicle;
Using a sensor model that models the sensors of the vehicle, creating sensor data that will be acquired when the vehicle runs the scenario set by the scenario setting unit;
A step of associating the created sensor data with actual data that is data acquired by a sensor of the vehicle, converting the sensor data based on the associated actual data, and creating converted actual data that is evaluation data. and an evaluation data creation method including. An evaluation data creation program that creates evaluation data for evaluating at least one of a vehicle's automatic driving and a vehicle's driving support function, the program comprising:
setting a scenario that reproduces the driving of the vehicle;
creating sensor data to be obtained when the vehicle travels through a set scenario using a sensor model that models the sensors of the vehicle;
A step of associating the created sensor data with actual data that is data acquired by a sensor of the vehicle, converting the sensor data based on the associated actual data, and creating converted actual data that is evaluation data. An evaluation data creation program that executes processing including.

Images