Automatic driving algorithm training and verification method based on virtual simulation scene

Abstract

The invention provides an automatic driving algorithm training and verification method based on a virtual simulation scene, and belongs to the technical field of data processing, and the method specifically comprises the steps: determining a deviation recognition target in a target object through recognition deviation data in different test environments, taking a recognition image with the deviation recognition target in the recognition data as a matched modeling image, and carrying out the recognition of the recognition image; based on the distribution association data of the deviation identification target in the matched modeling image and the identification deviation data of the deviation identification target in the test scene, determining a matched modeling image subjected to amplification processing in the test scene, and determining a modeling amplification image in the test scene in combination with the matched modeling image; the modeling processing of the test data in the test scene is performed based on the modeling amplification image, the matching modeling image and the identification deviation image in the test scene, and the training and verification processing of the automatic driving algorithm are performed by using the test data, so that the training processing efficiency and accuracy are improved.

Description

Technical Field

[0001] The present invention belongs to the field of data processing technology, and in particular relates to an automatic driving algorithm training and verification method based on virtual simulation scenarios. Background Art

[0002] In order to realize the training and processing of the autonomous driving algorithm, the existing technical solutions often carry out the training and verification of the autonomous driving algorithm in the real environment. However, there are many defects in the construction of the real driving scene. In order to realize the training and processing of the autonomous driving algorithm, the invention patent application CN202411162675.6 "A method, system and medium for generating autonomous driving simulation test scenes" imports the test scene into the simulation platform for testing, so that the test scene containing snowy features can be automatically generated, which fully covers various situations that may be encountered in driving on snowy days. The generated snowy simulation test scene can be used to test the safety of the autonomous driving system. However, there are the following technical problems: When constructing a virtual scene, since there are large differences in the recognition accuracy of different objects in different driving scenarios, if the virtual scene cannot be constructed based on the recognition accuracy, the reliability of the training process cannot be guaranteed.

[0003] In response to the above technical problems, the present invention provides an autonomous driving algorithm training and verification method based on virtual simulation scenarios. Summary of the invention

[0004] To achieve the purpose of the present invention, the present invention adopts the following technical solutions: According to one aspect of the present invention, a method for training and verifying an autonomous driving algorithm based on a virtual simulation scenario is provided.

[0005] A method for training and verifying an autonomous driving algorithm based on a virtual simulation scenario, specifically comprising: S1 determines recognition deviation data of target objects in different test scenarios using recognition data of the autonomous driving algorithm in different test scenarios, and proceeds to the next step when the recognition deviation state in the test scenario is determined to meet the requirements using the recognition deviation data; S2 determines recognition deviation data of different target objects in different test scenarios based on recognition data of different target objects in different test scenarios, and determines deviation recognition targets in the target objects with the recognition deviation data in different test environments; S3 uses the recognition image of the deviation recognition target in the recognition data as the matching modeling image, determines the matching modeling image for augmentation processing in the test scene based on the distribution association data of the deviation recognition target in the matching modeling image and the recognition deviation data of the deviation recognition target in the test scene, and determines the modeling augmented image in the test scene in combination with the matching modeling image; S4 performs modeling processing on the test data in the test scene based on the modeling augmented image, the matching modeling image, and the identification deviation image in the test scene, and uses the test data to perform training and verification processing on the autonomous driving algorithm.

[0006] A further technical solution is that the test scenes include rain, cloudy days, snow, and heavy fog.

[0007] A further technical solution is that the recognition deviation data of the target object includes the number of recognition deviations of the target object.

[0008] A further technical solution is to determine whether the recognition deviation state in the test scenario meets the requirements, specifically including: Determine the recognition deviation times of different target objects in the test scene using the recognition deviation data, and determine the recognition deviation objects among the target objects using the recognition deviation times; It is determined whether the recognition deviation state in the test scene meets the requirement according to the number of the recognition deviation objects.

[0009] A further technical solution is that the identification deviation object is a target object whose identification deviation times are greater than a preset deviation times.

[0010] A further technical solution is that the enlargement processing enlarges and reduces the image size of the identification deviation target in the matching modeling image according to a preset level.

[0011] A further technical solution is to perform modeling processing on the test data in the test scenario, specifically including: The modeling augmented image, the matching modeling image, and the identification deviation image in the test scene are used as basic data, and based on the basic data, modeling processing of the test data in the test scene is performed.

[0012] A further technical solution is to use test data from different test scenarios to train and verify the autonomous driving algorithm, specifically including: The model constructed by the autonomous driving algorithm is imported into the virtual simulation scene constructed by different test scenarios. The operation mode is selected, and the virtual car automatically runs in the scene to complete the road condition recognition in the road scene. According to the recognition results, the car can make corresponding actions to complete the autonomous driving algorithm training and verification of the car in the virtual simulation scene.

[0013] The beneficial effects of the present invention are: Based on the distribution association data of the deviation recognition targets in the matching modeling image and the recognition deviation data of the deviation recognition targets in the test scene, the matching modeling image for augmented processing in the test scene is determined, thereby fully considering the distribution of different recognition deviation targets in the actual recognition scene, and realizing the screening of matching modeling images with a large distribution of recognition deviation targets or recognition deviations that do not meet the requirements, which also lays the foundation for improving the test reliability in the test scene.

[0014] The test data in the test scene are modeled based on the modeling augmented images, matching modeling images and recognition deviation images in the test scene. Not only the recognition deviation images in the test scene are taken into account, but also the matching modeling images containing deviation recognition targets and the modeling augmented images with a higher probability of recognition deviation are further considered. This enables the modeling processing of the test data in the test scene from multiple angles, thereby improving the reliability of the training and verification processing of the autonomous driving algorithm.

[0015] Other features and advantages will be described in the following description. The objects and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description and drawings.

[0016] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The above and other features and advantages of the present invention will become more apparent by describing in detail exemplary embodiments thereof with reference to the attached drawings.

[0018] Figure 1 It is a flowchart of an autonomous driving algorithm training and verification method based on virtual simulation scenarios; Figure 2 is a flow chart of a method for determining that a recognition deviation state in a test scenario meets requirements; Figure 3 is a flow chart of a method for determining a deviation identification target in a target object; Figure 4 is a flow chart of a method for determining a matching modeled image for augmentation processing in a test scene. DETAILED DESCRIPTION

[0019] When constructing a virtual scene, the present invention constructs the virtual scene according to the recognition accuracy of different objects of the vehicle in different driving scenes, thereby ensuring the recognition accuracy and reducing the difficulty of constructing the virtual scene.

[0020] Determine whether the recognition deviation state in the test scene meets the requirements: determine the number of recognition deviations of the target object in different test scenes, and when the number of recognition deviations in the test scene is greater than 100 times, determine that the recognition deviation state in the test scene does not meet the requirements.

[0021] Deviation recognition target: It is determined by the recognition deviation data of the target object in different test environments. Specifically, the number of test environments with recognition deviation times greater than 50 times is determined. When the number of test environments with recognition deviation times greater than 50 times is greater than 5, the recognition target is determined to be the recognition deviation target.

[0022] Matching modeling images for augmentation processing in the test scene: Matching modeling images with a total number of recognition deviations of the deviation recognition target greater than 200 times are used as matching modeling images for augmentation processing in the test scene.

[0023] Perform modeling processing on the test data in the test scene: use the modeling augmented image, matching modeling image, and recognition deviation image in the test scene as basic data, and based on the basic data, perform modeling processing on the test data in the test scene. Import the model constructed by the autonomous driving algorithm into the virtual simulation scene constructed by different test scenes, select the operation mode, and let the virtual car automatically run in the scene to complete the road condition recognition in the road scene. According to the recognition result, the car can make corresponding actions to complete the autonomous driving algorithm training and verification of the car in the virtual simulation scene.

[0024] Driving data collection Video Recording: Enter the recording mode in the virtual platform to collect data. In this mode, the user will move in the scene in the first person and can modify the scene parameters (weather, traffic lights, road signs, etc.) through the function menu, so as to customize the environment and traffic signs of the entire virtual scene and collect data from multiple angles. Click Start Recording to start recording the current screen, and click Stop Recording to automatically save this screen in the specified directory in MP4 format.

[0025] Script processing: Read multiple mp4 files and save the images in the file to the specified folder according to the read frame frequency (if the current image is the same as the last saved image, it will not be saved repeatedly to ensure the diversity of training materials).

[0026] Data labeling: Process the video into labelable images and then label the data.

[0027] Annotation implementation algorithm: When the user clicks on the chart list, the image is rendered to the specified area.

[0028] When the user starts to annotate, the x-axis and y-axis distances of the clicked area from the upper left corner of the window (origin) and the x-axis and y-axis distances of the image from the origin are first judged: if the x-axis and y-axis distances of the clicked position are both smaller than the distance between the upper left corner of the image and the origin, it is judged as an external click, and the annotation box is annotated starting from the upper left corner of the image; if the current x-axis or y-axis position from the origin exceeds the distance from the image to the upper left corner of the window by the x-axis plus the image width or the y-axis plus the image length, it is judged as exceeding, and the annotation box position is specified as the maximum position corresponding to the image distance from the origin to prevent the annotation box from exceeding the image.

[0029] When the user starts to drag the mouse to mark, the above judgment and corresponding operations are performed in real time while dragging, and data is calculated and stored in real time. Four data need to be stored (the ratio of the distance point of the center of the annotation box on the x-axis in the picture divided by the width of the picture, the ratio of the distance point of the center of the annotation box on the y-axis in the picture divided by the height of the picture, the ratio of the width of the annotation box divided by the width of the picture, and the ratio of the height of the annotation box divided by the height of the picture). The width and height of the annotation box are the absolute values ​​of the distance moved by the mouse after the initial click position (a positive number for the mouse moving to the right, a negative number for the mouse moving to the left, a positive number for the mouse moving downward, and a negative number for the mouse moving upward). The position of the center point of the annotation box on the x-axis in the picture is the position of the initial click position on the x-axis of the picture plus the positive or negative number divided by 2, and the position of the center point of the annotation box on the y-axis in the picture is the position of the initial click position on the y-axis of the picture plus the positive or negative number divided by 2.

[0030] Dividing the above 4 values ​​by the width and height of the image will give the above 4 values ​​and store them.

[0031] When the user releases the mouse, the marking is considered complete; Save the annotation information to a file with the same name as the image and a txt suffix.

[0032] Model Training Arrange the annotated images and annotation information documents into corresponding pairs, randomly divide them into training set, test set, and validation set according to the specified division ratio, and automatically generate the configuration file required for YOLO training. Use YOLO's training script and start training after specifying the configuration file path Simulation Verification Use yolo's detect.py to convert the trained model into the universal model onnx format and save it to the specified path of the virtual platform. Start the virtual platform and enter the running mode. In this mode, the virtual platform will load the scene (weather and road signs of different sections of the road) through the pre-set scene configuration, and load the model under the specified path for local reasoning.

[0033] After the model is successfully loaded, the car moves forward naturally, and the pictures captured by the camera (from the car's perspective, located at the front of the car) are transmitted to the model for inference, and the frame number of the frame and the camera information (position, rotation, FOV, etc.) are marked.

[0034] Result display and scoring: Control the car's driving behavior in the scene (turning, accelerating / decelerating, turning on lights, honking, etc.) according to the model results. At the same time, according to the frame number corresponding to the result, obtain the camera information corresponding to the frame, try to find the corresponding object (road signs, traffic lights, pedestrians, etc.) recognized by the model by emitting rays, and compare it with the correct result corresponding to the object, and record the score. If the model reasoning result is wrong, the car drives into the wrong road or arrives at the wrong place, it will trigger the car's position backtracking to help the car return to the correct road (the scoring system will not give points if the reasoning is wrong or the comparison result is wrong).

[0035] To solve the above problems, according to one aspect of the present invention, Figure 1 As shown, a method for training and verifying an autonomous driving algorithm based on a virtual simulation scenario is provided, which specifically includes: S1 determines recognition deviation data of target objects in different test scenarios using recognition data of the autonomous driving algorithm in different test scenarios, and proceeds to the next step when the recognition deviation state in the test scenario is determined to meet the requirements using the recognition deviation data; S2 determines recognition deviation data of different target objects in different test scenarios based on recognition data of different target objects in different test scenarios, and determines deviation recognition targets in the target objects with the recognition deviation data in different test environments; S3 uses the recognition image of the deviation recognition target in the recognition data as the matching modeling image, determines the matching modeling image for augmentation processing in the test scene based on the distribution association data of the deviation recognition target in the matching modeling image and the recognition deviation data of the deviation recognition target in the test scene, and determines the modeling augmented image in the test scene in combination with the matching modeling image; S4 performs modeling processing on the test data in the test scene based on the modeling augmented image, the matching modeling image, and the identification deviation image in the test scene, and uses the test data to perform training and verification processing on the autonomous driving algorithm.

[0036] Furthermore, the test scenarios include rain, cloudy days, snow, and heavy fog.

[0037] Specifically, the recognition deviation data of the target object includes the number of recognition deviations of the target object.

[0038] It should be noted that if Figure 2 As shown, determining that the recognition deviation state in the test scenario meets the requirements specifically includes: Determine the recognition deviation times of different target objects in the test scene using the recognition deviation data, and determine the recognition deviation objects among the target objects using the recognition deviation times; It is determined whether the recognition deviation state in the test scene meets the requirement according to the number of the recognition deviation objects.

[0039] Furthermore, the identification deviation object is a target object whose identification deviation times are greater than a preset deviation times.

[0040] Optionally, determining whether the recognition deviation state in the test scene meets the requirement according to the number of the recognition deviation objects specifically includes: When the number of the recognition deviation objects is greater than the preset number of objects, it is determined that the recognition deviation state in the test scene does not meet the requirements.

[0041] It should also be noted that when the recognition deviation state in the test scene does not meet the requirements, all the recognition images in the test scene are used to perform modeling processing on the test data in the test scene.

[0042] Optionally, determining whether the recognition deviation state in the test scenario meets the requirement specifically includes: Determine the recognition deviation conditions of different recognition images in the test scene using the recognition deviation data, and determine the recognition deviation image in the recognition image according to the recognition deviation conditions; Whether the recognition deviation state in the test scene meets the requirement is determined according to the number of the recognition deviation images.

[0043] In another embodiment, determining that the recognition deviation state in the test scenario meets the requirement specifically includes: S11 determines the recognition deviation of different recognition images in the test scene using the recognition deviation data, determines the recognition deviation image in the recognition image according to the recognition deviation, and determines the basic deviation coefficient using the proportion of the recognition deviation image in the recognition image; Optionally, the above step S11 includes the following contents: S111 uses the recognition deviation data to determine the recognition deviation conditions of different recognition images in the test scene, and determines the recognition deviation images in the recognition images according to the recognition deviation conditions. When the number of the recognition deviation images does not meet the requirement, it is determined that the recognition deviation state in the test scene does not meet the requirement. When the number of the recognition deviation images meets the requirement, the process proceeds to step S112. S112 determines a basic deviation coefficient using the proportion of the recognition deviation image in the recognition image. When the basic deviation coefficient does not meet the requirement, the process proceeds to step S113. When the basic deviation coefficient meets the requirement, the process proceeds to step S12. S113 obtains the number of the recognition deviation images. When the number of the recognition deviation images is within the preset image number range, it is determined that the recognition deviation state in the test scene does not meet the requirements. When the number of the recognition deviation images is not within the preset image number range, proceed to step S12.

[0044] S12: determining the recognition deviation times of different target objects in the test scene using the recognition deviation data, and determining the recognition deviation coefficients of different target objects using the recognition deviation times of different target objects; Optionally, the above step S12 includes the following contents: S121 determines the number of recognition deviations of different target objects in the test scene using the recognition deviation data. When the total number of recognition deviations of different target objects does not meet the requirement, it is determined that the recognition deviation state in the test scene does not meet the requirement. When the total number of recognition deviations of different target objects meets the requirement, the process proceeds to step S122. S122: When it is determined that there is a target object whose identification deviation number is greater than the preset deviation number based on the identification deviation numbers of different target objects, the process proceeds to step S123; when there is no target object whose identification deviation number is greater than the preset deviation number, the process proceeds to step S124; S123: When the number of target objects whose identification deviation times are greater than the preset deviation times does not meet the requirement, it is determined that the identification deviation state in the test scene does not meet the requirement; when the number of target objects whose identification deviation times are greater than the preset deviation times meets the requirement, the process proceeds to step S124; S124 determines the recognition deviation coefficients of different target objects using the recognition deviation times of different target objects. When the number of target objects whose recognition deviation coefficients are within the preset deviation coefficient interval is greater than the preset number of target objects, it is determined that the recognition deviation state in the test scene does not meet the requirements. When the number of target objects whose recognition deviation coefficients are within the preset deviation coefficient interval is not greater than the preset number of target objects, proceed to step S13.

[0045] S13 determines a comprehensive deviation coefficient according to the recognition deviation coefficients of different target objects and the basic deviation coefficient, and uses the comprehensive deviation coefficient to determine whether the recognition deviation state in the test scene meets the requirements.

[0046] Specifically, Figure 3 As shown, the method for determining the deviation identification target in the target object is: Determine the number of recognition deviations of the target object in the different test environments using the recognition deviation data of the target object in the different test environments; Determining a total number of recognition deviations of the target object based on the number of recognition deviations of the target object in different test environments; Whether the target object is a deviation identification target is determined according to the total number of identification deviations.

[0047] Furthermore, when the total number of identification deviations of the target object is greater than a preset deviation number threshold, the target object is determined to be a deviation identification target.

[0048] Optionally, the method for determining the deviation identification target in the target object is: Determine the number of recognition deviations of the target object in the different test environments using the recognition deviation data of the target object in the different test environments; Determining recognition deviation environments of the target object in different test environments based on the number of recognition deviations of the target object in different test environments; Whether the target object is a deviation recognition target is determined according to the number of the recognition deviation environments.

[0049] In another embodiment, the method for determining the deviation identification target in the target object is: S21 determines the number of recognition deviations of the target object in different test environments based on the recognition deviation data of the target object in different test environments; S22 determining recognition deviation coefficients of the target object in different test environments based on the number of recognition deviations of the target object in different test environments and the number of recognition images containing the target object; S23 determines the comprehensive recognition deviation coefficient of the target object according to the recognition deviation coefficients of the target object in different test environments, and uses the comprehensive recognition deviation coefficient to determine whether the target object is a deviation recognition target.

[0050] Specifically, Figure 4 As shown, the method for determining the matching modeling image for augmentation processing in the test scene is: Determine the deviation recognition targets in different matching modeling images by using the distribution association data of the deviation recognition targets in the matching modeling images, and determine the number of recognition deviations of different recognition deviation targets in the test environment according to the recognition error data of the deviation recognition targets in the different matching modeling images in the test environment; Determine the image recognition deviation coefficient of the matching modeling image by using the recognition deviation times of different recognition deviation targets in the test environment; The matching modeling image for augmentation processing in the test scene is determined based on the image recognition deviation coefficient.

[0051] Furthermore, the matching modeling image that is augmented in the test scene is a matching modeling image whose image recognition deviation coefficient does not meet the requirements.

[0052] Optionally, the method for determining the matching modeling image for augmentation processing in the test scene is: Determine the deviation recognition targets in different matching modeling images by using the distribution association data of the deviation recognition targets in the matching modeling images, and determine the number of recognition deviations of different recognition deviation targets in the test environment according to the recognition error data of the deviation recognition targets in the different matching modeling images in the test environment; Determine the image recognition deviation coefficient of the matching modeling image by using the recognition deviation times of different recognition deviation targets in the test environment; The matching modeling image for augmentation processing in the test scene is determined based on the image recognition deviation coefficient.

[0053] It should be noted that the matching modeling image that is augmented in the test scene is a matching modeling image whose image recognition deviation coefficient does not meet the requirements.

[0054] Optionally, the method for determining the matching modeling image for augmentation processing in the test scene is: Determine the deviation recognition targets in different matching modeling images by using the distribution association data of the deviation recognition targets in the matching modeling images, and determine the number of recognition deviations of different recognition deviation targets in the test environment according to the recognition error data of the deviation recognition targets in the different matching modeling images in the test environment; When there is a recognition deviation number in the matching modeling image that does not meet the required recognition deviation target, the matching modeling image belongs to the matching modeling image processed by augmentation in the test scene; When there is no recognition deviation number in the matching modeling image that does not meet the required recognition deviation target: When it is determined that the identification deviation times of different identification deviation targets in the test environment are all less than a preset deviation times threshold value: When the number of identified deviation targets in the matching modeling image is less than the preset number of deviation targets: It is determined that the matching model image does not belong to the matching model image augmented in the test scene; When the recognition deviation times of different recognition deviation targets are not all less than the preset deviation times threshold or the number of recognition deviation targets in the matching modeling image is not less than the preset deviation target number: Determining the total number of recognition deviations according to the number of recognition deviations of different recognition deviation targets in the test environment, and when the total number of recognition deviations does not meet the requirement, determining that the matching modeling image does not belong to the matching modeling image augmented in the test scene; When the total number of identification deviations meets the requirements: The identification deviation targets whose identification deviation times are within the preset deviation times interval are used as screening deviation targets, and when the number of screening deviation targets does not meet the requirement, it is determined that the matching modeling image does not belong to the matching modeling image augmented in the test scene; When the number of screening deviation targets meets the requirement: The image recognition deviation coefficient of the matching modeling image is determined by utilizing the recognition deviation times of different recognition deviation targets in the test environment, and the matching modeling image for augmentation processing in the test scene is determined based on the image recognition deviation coefficient.

[0055] Furthermore, the enlargement process enlarges and reduces the image size of the identification deviation target in the matching modeling image according to a preset level.

[0056] Specifically, the modeling process of the test data in the test scenario includes: The modeling augmented image, the matching modeling image, and the identification deviation image in the test scene are used as basic data, and based on the basic data, modeling processing of the test data in the test scene is performed.

[0057] It should be noted that the training and verification of the autonomous driving algorithm using test data in different test scenarios specifically includes: The model constructed by the autonomous driving algorithm is imported into the virtual simulation scene constructed by different test scenarios. The operation mode is selected, and the virtual car automatically runs in the scene to complete the road condition recognition in the road scene. According to the recognition results, the car can make corresponding actions to complete the autonomous driving algorithm training and verification of the car in the virtual simulation scene.

[0058] Optionally, the above step S21 includes the following contents: S211 determines the number of recognition deviations of the target objects in different test environments based on the recognition deviation data of the target objects in different test environments. When the total number of recognition deviations of the target objects in different test environments does not meet the requirement, the target object is determined to be a recognition deviation target. When the total number of recognition deviations of the target objects in different test environments meets the requirement, the process proceeds to step S212. When it is determined in step S212 that there is a recognition deviation test environment in the test environment by the number of recognition deviations of the target object in different test environments, the process proceeds to step S213; when there is no recognition deviation test environment in the test environment, the process proceeds to step S22; S213: When the number of the recognition deviation test environments does not meet the requirement, the target object is determined to be a recognition deviation target. When the number of the recognition deviation test environments meets the requirement, the process proceeds to step S22.

[0059] Optionally, the above step S22 includes the following contents: S221 determines the recognition deviation coefficient of the target object in different test environments based on the number of recognition deviations of the target object in different test environments and the number of recognition images containing the target object. When there is a test environment where the recognition deviation coefficient does not meet the requirements, the process proceeds to step S222. When there is no test environment where the recognition deviation coefficient does not meet the requirements, the process proceeds to step S23. S222: taking the test environment whose recognition deviation coefficient does not meet the requirement as the recognition deviation test environment; when the number of the recognition deviation test environment does not meet the requirement, determining the target object as the recognition deviation target; when the number of the recognition deviation test environment meets the requirement, proceeding to step S223; S223 determines the environment recognition deviation coefficient with the recognition deviation coefficient of different recognition deviation test environments. When the environment recognition deviation coefficient does not meet the requirements, the target object is determined to be a recognition deviation target. When the environment recognition deviation coefficient meets the requirements, proceed to step S23.

[0060] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device, equipment, and non-volatile computer storage medium embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0061] The above is a description of a specific embodiment of the specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0062] The above description is only one or more embodiments of this specification and is not intended to limit this specification. For those skilled in the art, one or more embodiments of this specification may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of one or more embodiments of this specification shall be included in the scope of the claims of this specification.

Claims

1. A method for training and verifying an autonomous driving algorithm based on a virtual simulation scenario, characterized in that: Specifically include: Determine recognition deviation data of target objects in different test scenarios using recognition data of the autonomous driving algorithm in different test scenarios, and proceed to the next step when determining using the recognition deviation data that the recognition deviation state in the test scenario meets the requirements; Determine recognition deviation data of different target objects in different test scenarios based on recognition data of different target objects in different test scenarios, and determine deviation recognition targets in the target objects with the recognition deviation data in different test environments; Taking the recognition image of the deviation recognition target in the recognition data as the matching modeling image, determining the matching modeling image of the augmentation processing in the test scene based on the distribution association data of the deviation recognition target in the matching modeling image and the recognition deviation data of the deviation recognition target in the test scene, and determining the modeling augmented image in the test scene in combination with the matching modeling image; Based on the modeling augmented image, the matching modeling image, and the identification deviation image in the test scene, the test data in the test scene is modeled, and the test data is used to train and verify the autonomous driving algorithm.

2. The method for training and verifying an autonomous driving algorithm based on a virtual simulation scenario according to claim 1, characterized in that: The test scenarios include rain, cloudy days, snow, and heavy fog.

3. The method for training and verifying an autonomous driving algorithm based on a virtual simulation scenario according to claim 1, characterized in that: The recognition deviation data of the target object includes the number of recognition deviations of the target object.

4. The method for training and verifying an autonomous driving algorithm based on a virtual simulation scenario according to claim 1, wherein: Determining that the recognition deviation state in the test scenario meets the requirements, specifically includes: Determine the recognition deviation times of different target objects in the test scene using the recognition deviation data, and determine the recognition deviation objects among the target objects using the recognition deviation times; It is determined whether the recognition deviation state in the test scene meets the requirement according to the number of the recognition deviation objects.

5. The method for training and verifying an autonomous driving algorithm based on a virtual simulation scenario according to claim 4, characterized in that: The recognition deviation object is a target object whose recognition deviation times are greater than a preset deviation times.

6. The method for training and verifying an autonomous driving algorithm based on a virtual simulation scenario according to claim 4, characterized in that: Determining whether the recognition deviation state in the test scene meets the requirements according to the number of the recognition deviation objects specifically includes: When the number of the recognition deviation objects is greater than the preset number of objects, it is determined that the recognition deviation state in the test scene does not meet the requirements.

7. The method for training and verifying an autonomous driving algorithm based on a virtual simulation scenario according to claim 1, characterized in that: The enlargement process enlarges and reduces the image size of the identification deviation target in the matching modeling image according to a preset level.

8. The method for training and verifying an autonomous driving algorithm based on a virtual simulation scenario according to claim 1, characterized in that: The modeling process of the test data in the test scenario specifically includes: The modeling augmented image, the matching modeling image, and the identification deviation image in the test scene are used as basic data, and based on the basic data, modeling processing of the test data in the test scene is performed.

9. The method for training and verifying an autonomous driving algorithm based on a virtual simulation scenario according to claim 1, wherein: The autonomous driving algorithm is trained and verified using test data from different test scenarios, specifically including: The model constructed by the autonomous driving algorithm is imported into the virtual simulation scene constructed by different test scenarios. The operation mode is selected, and the virtual car automatically runs in the scene to complete the road condition recognition in the road scene. According to the recognition results, the car can make corresponding actions to complete the autonomous driving algorithm training and verification of the car in the virtual simulation scene.

Images