A true value evaluation method and device based on full FOV annotation

Abstract

The present invention provides a true value evaluation method and device based on the full FOV annotation method, which solves the problem of FOV loss and poor compatibility between true value evaluation data sets of different projects. Through the data of different camera combinations and laser radar data collected in different projects, the true value data and the re-injection data are constructed, and by converting the true value data and the re-injection data to the same reference coordinate system, the quasi-recall evaluation of the same obstacle and the obstacle distance measurement, speed measurement and obstacle length, width and height evaluation results are achieved. The full FOV annotation of the autonomous driving scene pictures is achieved, and the compatibility of the evaluation data set in the field of autonomous driving is improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent automobile technology, and in particular to a true value evaluation method and device based on a full FOV annotation method. Background Art

[0002] With the rapid development of fields such as self-driving cars, drones, and robot vision systems, higher requirements are placed on the accuracy and comprehensiveness of environmental perception.

[0003] Field of View (FOV) annotation can provide rich training data, so that the model can learn to make correct decisions in complex and changing environments. For example, in autonomous driving, the model needs to identify all obstacles, pedestrians, traffic signs and road conditions on the road, which requires a large amount of FOV annotation data. Conventional distortion removal methods damage the camera perspective. This incomplete annotation method may lead to inaccurate evaluation of algorithm performance and fail to fully reflect the performance of the algorithm in actual application scenarios. At the same time, in the prior art, for different projects, the camera combinations used are mostly different, and the camera's FOV and image size may also be different. Therefore, every time a new project is started, it is necessary to start annotating a batch of new data for evaluation. The cost of building a data set for true value evaluation is relatively high, and after the project is completed, the subsequent utilization value of the true value evaluation data of the project is also relatively low, and the compatibility between the true value evaluation data sets of different projects is poor. Summary of the invention

[0004] The present invention provides a true value evaluation method and device based on a full FOV annotation method to solve the problems of FOV loss and poor compatibility between true value evaluation data sets of different projects.

[0005] In a first aspect, the present invention provides a true value evaluation method based on a full FOV annotation method, which specifically comprises the following steps:

[0006] Step S1, determining a project, obtaining corresponding camera group video data according to the project, and generating camera group picture data DATA according to the camera group video data;

[0007] Step S2, generating a true value result DATA1 according to the camera group picture data DATA;

[0008] Step S3, generating a refill result DATA2 according to the camera group picture data DATA;

[0009] Step S4, calculating the evaluation result according to the true value result DATA1 and the re-injection result DATA2.

[0010] Among them, the project types include one or at least two.

[0011] Preferably, in step S1, the data source of the camera group video data DATA includes a combination of one or more camera data of a front-view camera, a rear-view camera, a periscopic camera, a surround-view camera, a monocular camera, and a binocular camera. For example: front-view monocular camera data, front-view monocular camera data + periscopic camera data, front-view binocular camera data + periscopic camera data + rear-view camera data.

[0012] Preferably, in step S2, generating a true value result DATA1 according to the camera group picture data DATA specifically includes the following steps:

[0013] Step S201, obtaining a pixel width W of a viewing angle of any camera group image data DATA and a camera horizontal viewing angle FOV, and calculating a new focal length new_fx according to the pixel width W and the camera horizontal viewing angle FOV;

[0014] Step S202: Replace the new focal length new_fx into the original camera intrinsic parameter matrix to form a new intrinsic parameter matrix;

[0015] Step S203, converting the original image into a full FOV image after distortion removal through the original internal parameter matrix, the new internal parameter matrix and the distortion coefficient;

[0016] Step S204, repeatedly executing steps S201 to S203, obtaining a plurality of full FOV images of different viewing angles and marking the 2D information of obstacles on the full FOV images of different viewing angles;

[0017] Step S205: Obtain laser point cloud through laser radar, and mark 3D information on the laser point cloud;

[0018] Step S206: Fuse the 2D information and the 3D information to generate a true value result DATA1.

[0019] Preferably, in step S204, the 2D information includes box information, obstacle category and other information.

[0020] Preferably, in step S205, the 3D information includes distance, length, width and height information.

[0021] Preferably, in step S3, generating a refill result DATA2 according to the camera group picture data DATA specifically includes the following steps:

[0022] Step S301, performing a distortion removal process on the camera group image data DATA to generate the distortion-removed data data-1;

[0023] Step S302, cropping and scaling the data data-1 to generate data data-2;

[0024] Step S303: The data data-2 is processed through a model and post-processing to generate a reinjection result DATA2.

[0025] Preferably, in step S301, the camera group image data DATA is subjected to distortion removal processing by using the original camera internal parameter matrix, the distortion coefficient and the virtual internal parameter matrix set by the project to form the distortion-removed data data-1.

[0026] Preferably, in step S303, the post-processing specifically includes calculating information such as distance measurement and speed measurement according to the result of the obstacle recognition detection model (used to recognize obstacles in the picture), and determining whether it is the same obstacle under different camera viewing angles.

[0027] Preferably, in step S4, calculating the evaluation result according to the true value result DATA1 and the re-injection result DATA2 specifically includes the following steps:

[0028] Step S401, performing reverse scaling and cropping processing on the refill result DATA2 to generate 2D box data A based on the refill image reference coordinate system of the original size;

[0029] Step S402: According to the 2D information of the project camera, the 2D box of the true value result DATA1 is converted to the reference coordinate system of the re-injected image, and the 2D box data B of the true value result DATA1 in the reference coordinate system of the re-injected image is generated;

[0030] Step S403: perform IoU matching on data A and data B to calculate the evaluation result.

[0031] The evaluation results include obstacle recall evaluation results, obstacle distance measurement results, obstacle speed measurement results, and obstacle length, width, height and other accuracy evaluation results.

[0032] Preferably, in step S402, the 2D box of the true value result DATA1 is converted to the re-injected picture reference coordinate system through the virtual internal parameter matrix and the new internal parameter matrix.

[0033] In a second aspect, the present invention further provides a truth value evaluation device based on a full FOV annotation method, which specifically includes the following modules:

[0034] A project data acquisition module, used to determine a project, acquire corresponding camera group video data according to the project, and generate camera group picture data DATA according to the camera group video data;

[0035] A true value result generating module, used for generating a true value result DATA1 according to the camera group picture data DATA;

[0036] A refill result generating module, used for generating a refill result DATA2 according to the camera group picture data DATA;

[0037] The true value evaluation module is used to calculate the evaluation result according to the true value result DATA1 and the re-injection result DATA2.

[0038] Among them, the project types include one or at least two.

[0039] Preferably, in the project data acquisition module, the data source of the camera group video data DATA includes a combination of one or more camera data of a front-view camera, a rear-view camera, a periscopic camera, a surround-view camera, a monocular camera, and a binocular camera. For example: front-view monocular camera data, front-view monocular camera data + periscopic camera data, front-view binocular camera data + periscopic camera data + rear-view camera data.

[0040] Preferably, the true value result generation module specifically includes the following submodules:

[0041] The first submodule of the true value result generation is used to obtain the pixel width W of a viewing angle of any camera group image data DATA and the camera horizontal viewing angle FOV, and calculate the new focal length new_fx according to the pixel width W and the camera horizontal viewing angle FOV;

[0042] The true value result generates a second submodule, which is used to replace the new focal length new_fx into the original camera intrinsic parameter matrix to form a new intrinsic parameter matrix;

[0043] The third submodule of the true value result generation is used to convert the original image into a dedistorted full FOV image through the original internal parameter matrix, the new internal parameter matrix and the distortion coefficient;

[0044] A fourth submodule for generating a true value result is used to repeatedly execute the first submodule for generating a true value result, the second submodule for generating a true value result, and the third submodule for generating a true value result, to obtain a plurality of full FOV images of different viewing angles and to mark 2D information of obstacles on the full FOV images of different viewing angles;

[0045] The fifth submodule of the true value result generation is used to obtain the laser point cloud through the laser radar and mark the 3D information on the laser point cloud;

[0046] The sixth submodule of true value result generation is used to fuse the 2D information and the 3D information to generate a true value result DATA1.

[0047] Preferably, in the fourth submodule of generating the true value result, the 2D information includes box information, obstacle category and other information.

[0048] Preferably, in the fifth submodule of generating the true value result, the 3D information includes information such as distance and length, width and height.

[0049] Preferably, the recharging result generating module specifically includes the following submodules:

[0050] The first submodule for generating the feedback result is used to perform a distortion removal process on the camera group image data DATA to generate the distortion-removed data data-1;

[0051] The second submodule of the refill result generation is used to crop and scale the data data-1 to generate data data-2;

[0052] The third submodule of the reinjection result generation is used to generate the reinjection result DATA2 by passing the data data-2 through the model and post-processing.

[0053] Preferably, in the first submodule for generating the re-injection result, the camera group image data DATA is subjected to distortion removal processing by using the original camera internal parameter matrix, the distortion coefficient and the virtual internal parameter matrix set by the project to form the distortion-removed data data-1.

[0054] Preferably, in the third submodule for generating the feedback result, the post-processing specifically includes calculating information such as distance measurement and speed measurement according to the result of the obstacle recognition detection model (used to identify obstacles in the picture), and determining whether it is the same obstacle under different camera viewing angles.

[0055] Preferably, the truth value evaluation module specifically includes the following submodules:

[0056] The first submodule of the true value evaluation is used to reversely scale and crop the refill result DATA2 to generate 2D box data A in the reference coordinate system of the refilled image based on the original size;

[0057] The second submodule of the true value evaluation is used to convert the 2D box of the true value result DATA1 to the reference coordinate system of the re-injected image according to the 2D information of the project camera, and generate the 2D box data B of the true value result DATA1 in the reference coordinate system of the re-injected image;

[0058] The third submodule of the true value evaluation is used to perform IoU matching between data A and data B and calculate the evaluation results.

[0059] The evaluation results include obstacle recall evaluation results, obstacle distance measurement results, obstacle speed measurement results, and obstacle length, width, height and other accuracy evaluation results.

[0060] Preferably, in the second submodule of the truth value evaluation, the 2D box of the truth value result DATA1 is converted to the reference coordinate system of the re-injected image through the virtual internal parameter matrix and the new internal parameter matrix.

[0061] In a third aspect, the present invention further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a truth value evaluation method based on a full FOV annotation method as described in any one of the first aspects of the present application.

[0062] In a fourth aspect, the present invention further provides an electronic device, comprising: a memory storing a computer program; and a processor communicatively connected to the memory, for executing a truth value evaluation method based on a full FOV annotation method as described in any one of the first aspects of the present application when the computer program is called.

[0063] Compared with the prior art, the present invention has the following obvious outstanding substantive features and significant advantages:

[0064] The present invention provides a true value evaluation method and device based on the full FOV annotation method, which solves the problem of FOV loss and poor compatibility between true value evaluation data sets of different projects. Through the data of different camera combinations and laser radar data collected in different projects, the true value data and the re-injection data are constructed, and by converting the true value data and the re-injection data to the same reference coordinate system, the quasi-recall evaluation of the same obstacle and the obstacle distance measurement, speed measurement and obstacle length, width and height evaluation results are achieved. The full FOV annotation of the autonomous driving scene pictures is achieved, and the compatibility of the evaluation data set in the field of autonomous driving is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] The accompanying drawings, which constitute part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention. In the accompanying drawings:

[0066] Figure 1 It is a flow chart of a true value evaluation method based on a full FOV annotation method according to a preferred embodiment of the present invention.

[0067] Figure 2 It is a schematic diagram of a truth value evaluation method based on a full FOV annotation method according to a preferred embodiment of the present invention.

[0068] Figure 3 It is the original picture of the viewing angle V of the preferred embodiment of the present invention.

[0069] Figure 4 It is a partial enlarged view of the original image at the viewing angle V of the preferred embodiment of the present invention.

[0070] Figure 5 It is a FOV picture of the viewing angle V of the preferred embodiment of the present invention.

[0071] Figure 6 It is a partial enlarged view of the FOV picture of the viewing angle V of the preferred embodiment of the present invention.

[0072] Figure 7 It is a refill picture of the viewing angle V of the preferred embodiment of the present invention.

[0073] Figure 8 It is a partial enlarged view of the FOV picture of the viewing angle V of the preferred embodiment of the present invention.

[0074] Fig. 9 It is a structural schematic diagram of a truth value evaluation device based on a full FOV annotation method according to a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0075] The present invention provides a true value evaluation method and device based on a full FOV annotation method. In order to make the purpose, technical solution and effect of the present invention clearer and more specific, the present invention is further described in detail with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0076] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0077] Embodiment 1:

[0078] like Figure 1-Figure 8 As shown, the true value evaluation method based on the full FOV annotation method described in this embodiment specifically includes the following steps (such as Figure 1-Figure 2 shown):

[0079] Step S1, determine the project, obtain the corresponding camera group video data according to the project, and generate camera group image data DATA according to the camera group video data. The project type includes one or at least two. The data source of the camera group video data DATA includes a combination of one or more camera data of a front-view camera, a rear-view camera, a periscopic camera, a surround-view camera, a monocular camera, and a binocular camera. For example: a combination of front-view monocular camera data, a front-view monocular camera data + periscopic camera data, a front-view binocular camera data + periscopic camera data + rear-view camera data, and other data combinations.

[0080] Step S2: Generate a true value result DATA1 according to the camera group image data DATA.

[0081] Optionally, step S2 specifically includes the following steps:

[0082] Step S201, obtaining a pixel width W of a viewing angle of any camera group image data DATA and a camera horizontal viewing angle FOV, and calculating a new focal length new_fx according to the pixel width W and the camera horizontal viewing angle FOV.

[0083] Step S202: Replace the new focal length new_fx into the original camera intrinsic parameter matrix to form a new intrinsic parameter matrix.

[0084] Step S203: transform the original image (such as Figure 3 , Figure 4 ) is converted to a dedistorted full FOV image (as shown in Figure 5 , Figure 6 as shown).

[0085] Step S204, repeatedly executing steps S201 to S203, obtaining a plurality of full FOV images of different perspectives and marking the 2D information of obstacles on the full FOV images of different perspectives; wherein the 2D information includes box information, obstacle category and other information.

[0086] Step S205: Obtain laser point cloud through laser radar, and mark 3D information on the laser point cloud; wherein the 3D information includes distance, length, width, height and other information.

[0087] Step S206: Fusing the 2D information and the 3D information (ie, projecting the point cloud onto the image plane and fusing it with the 2D information obtained through manual adjustment and supplementation) to generate a true value result DATA1.

[0088] Step S3: Generate a refill result DATA2 according to the camera group picture data DATA.

[0089] Optionally, step S3 specifically includes the following steps:

[0090] Step S301, dedistorting the camera group image data DATA to generate dedistorted data data-1; wherein, dedistorting the camera group image data DATA is processed by using the original camera internal parameter matrix, distortion coefficients and a virtual internal parameter matrix set by the project to form the dedistorted data data-1.

[0091] Step S302: crop and scale the data data-1 to generate data data-2.

[0092] Step S303, the data data-2 is processed by the model and post-processing to generate the reinjection result DATA2, such as Figure 7-Figure 8 shown. Figure 7 , Figure 8 In the current frame, only one object’s error (rate) does not meet the KPI requirement. Figure 5 The selected picture is the visualization picture of the longitudinal distance error, such as Figure 5 , Figure 6 As shown in the object box, its true value (the part selected by the green box) is 82.66m, the predicted value (the part selected by the red box) is 76.32m, the error is 6.34m (the error represents the absolute value of the difference between the true value and the predicted value), and the KPI is 2.7m (KPI represents the maximum error allowed by the project); wherein, the post-processing specifically includes calculating the distance measurement, speed measurement and other information according to the results of the obstacle recognition detection model (used to identify obstacles in the picture), and judging whether it is the same obstacle under different camera perspectives.

[0093] Step S4, calculating the evaluation result according to the true value result DATA1 and the re-injection result DATA2.

[0094] Optionally, step S4 specifically includes the following steps:

[0095] Step S401: perform reverse scaling and cropping on the refill result DATA2 to generate 2D box data A based on the refill image reference coordinate system of the original size.

[0096] Step S402: According to the 2D information of the project camera, the 2D box of the true value result DATA1 is converted to the reference coordinate system of the re-injected image, and the 2D box data B of the true value result DATA1 in the reference coordinate system of the re-injected image is generated; wherein the 2D box of the true value result DATA1 is converted to the reference coordinate system of the re-injected image by means of the virtual intrinsic parameter matrix and the new intrinsic parameter matrix.

[0097] Step S403: perform IoU matching on data A and data B to calculate the evaluation result.

[0098] The evaluation results include obstacle recall evaluation results, obstacle distance measurement results, obstacle speed measurement results, and obstacle length, width, height and other accuracy evaluation results.

[0099] Embodiment 2:

[0100] like Fig. 9 As shown, the true value evaluation device based on the full FOV annotation method described in this embodiment specifically includes the following modules:

[0101] The project data acquisition module is used to determine the project, acquire the corresponding camera group video data according to the project, and generate camera group picture data DATA according to the camera group video data; wherein the project type includes one or at least two.

[0102] The data source of the camera group video data DATA includes a combination of one or more camera data of multiple front-view cameras, rear-view cameras, periscopic cameras, surround-view cameras, monocular cameras, and binocular cameras. For example: a combination of front-view monocular camera data, front-view monocular camera data + periscopic camera data, front-view binocular camera data + periscopic camera data + rear-view camera data, and the like.

[0103] The true value result generating module is used to generate a true value result DATA1 according to the camera group image data DATA.

[0104] The true value result generation module specifically includes the following submodules:

[0105] The first submodule of the true value result generation is used to obtain the pixel width W of a viewing angle of any camera group image data DATA and the camera horizontal viewing angle FOV, and calculate the new focal length new_fx according to the pixel width W and the camera horizontal viewing angle FOV.

[0106] The true value result generates a second submodule, which is used to replace the new focal length new_fx into the original camera intrinsic parameter matrix to form a new intrinsic parameter matrix.

[0107] The third submodule of the true value result generation is used to convert the original image into a dedistorted full FOV image through the original internal parameter matrix, the new internal parameter matrix and the distortion coefficient.

[0108] The fourth submodule of the true value result generation is used to repeatedly execute the first submodule of the true value result generation, the second submodule of the true value result generation and the third submodule of the true value result generation, obtain a plurality of full FOV images of different perspectives and mark the 2D information of the obstacles on the full FOV images of different perspectives; wherein the 2D information includes box information, obstacle category and other information.

[0109] The fifth submodule of the true value result generation is used to obtain the laser point cloud through the laser radar and mark the 3D information on the laser point cloud; wherein the 3D information includes distance and length, width and height information.

[0110] The sixth submodule of true value result generation is used to fuse the 2D information and the 3D information to generate a true value result DATA1.

[0111] The refill result generating module is used to generate a refill result DATA2 according to the camera group picture data DATA.

[0112] The recharging result generation module specifically includes the following submodules:

[0113] The first submodule of the feedback result generation is used to perform distortion removal processing on the camera group image data DATA to generate the distortion-removed data data-1; wherein, the camera group image data DATA is subjected to distortion removal processing by means of the original camera intrinsic parameter matrix, the distortion coefficient and the virtual intrinsic parameter matrix set by the project to form the distortion-removed data data-1.

[0114] The second submodule of the feedback result generation is used to crop and scale the data data-1 to generate data data-2.

[0115] The third submodule of the reinjection result generation is used to generate the reinjection result DATA2 by passing the data data-2 through the model and post-processing.

[0116] The true value evaluation module is used to calculate the evaluation result according to the true value result DATA1 and the re-injection result DATA2.

[0117] The true value evaluation module specifically includes the following submodules:

[0118] The first submodule of the true value evaluation is used to reversely scale and crop the re-injection result DATA2 to generate 2D box data A in a reference coordinate system of the re-injection image based on the original size.

[0119] The second submodule of the true value evaluation is used to convert the 2D box of the true value result DATA1 to the reference coordinate system of the re-injected image according to the 2D information of the project camera, and generate the 2D box data B of the true value result DATA1 in the reference coordinate system of the re-injected image; wherein, the 2D box of the true value result DATA1 is converted to the reference coordinate system of the re-injected image through the virtual intrinsic parameter matrix and the new intrinsic parameter matrix.

[0120] The third submodule of the true value evaluation is used to perform IoU matching through data A and data B to calculate the evaluation results; wherein the evaluation results include the obstacle quasi-recall evaluation results, obstacle ranging results, obstacle speed measurement results and obstacle length, width and height and other precision evaluation results.

[0121] The specific embodiments of the present invention are described in detail above, but they are only examples, and the present invention is not limited to the specific embodiments described above. For those skilled in the art, any equivalent modifications and substitutions made to the present invention are also within the scope of the present invention. Therefore, the equalization changes and modifications made without departing from the spirit and scope of the present invention should be included in the scope of the present invention.

Claims

1. A method for true value evaluation based on full FOV annotation, characterized in that: The specific steps include: Step S1, determining a project, obtaining corresponding camera group video data according to the project, and generating camera group picture data DATA according to the camera group video data; Step S2, generating a true value result DATA1 according to the camera group picture data DATA; Step S3, generating a refill result DATA2 according to the camera group picture data DATA; Step S4, calculating the evaluation result according to the true value result DATA1 and the re-injection result DATA2; Wherein, in step S1, the data source of the camera group video data includes a combination of one or more camera data of multiple front-view cameras, rear-view cameras, periscopic cameras, surround-view cameras, monocular cameras, and binocular cameras; In step S2, a true value result DATA1 is generated according to the camera group image data DATA, which specifically includes the following steps: Step S201, obtaining a pixel width W of a viewing angle of any camera group image data DATA and a camera horizontal viewing angle FOV, and calculating a new focal length new_fx according to the pixel width W and the camera horizontal viewing angle FOV; Step S202: Replace the new focal length new_fx into the original camera intrinsic parameter matrix to form a new intrinsic parameter matrix; Step S203, converting the original image into a full FOV image after distortion removal through the original internal parameter matrix, the new internal parameter matrix and the distortion coefficient; Step S204, repeatedly executing steps S201 to S203, obtaining a plurality of full FOV images of different viewing angles and marking the 2D information of obstacles on the full FOV images of different viewing angles; Step S205: Obtain laser point cloud through laser radar, and mark 3D information on the laser point cloud; Step S206, fusing the 2D information and the 3D information to generate a true value result DATA1; The 2D information includes box information and obstacle information, and the 3D information includes distance and length, width and height; In step S3, a refill result DATA2 is generated according to the camera group picture data DATA, which specifically includes the following steps: Step S301, performing a distortion removal process on the camera group image data DATA to generate the distortion-removed data data-1; Step S302, cropping and scaling the data data-1 to generate data data-2; Step S303, data data-2 is processed by the model and post-processed to generate data DATA2; Wherein, in step S301, the camera group image data DATA is subjected to distortion removal processing by using the original camera internal parameter matrix, the distortion coefficient and the virtual internal parameter matrix set by the project to form the distortion-removed data data-1.

2. The method for truth value evaluation based on full FOV annotation according to claim 1, characterized in that: In step S4, the evaluation result is calculated according to the true value result DATA1 and the re-injection result DATA2, which specifically includes the following steps: Step S401, performing reverse scaling and cropping processing on the refill result DATA2 to generate 2D box data A based on the refill image reference coordinate system of the original size; Step S402: According to the 2D information of the project camera, the 2D box of the true value result DATA1 is converted to the reference coordinate system of the re-injected image, and the 2D box data B of the true value result DATA1 in the reference coordinate system of the re-injected image is generated; Step S403, perform IoU matching on data A and data B, and calculate the evaluation result; The evaluation results include obstacle recall evaluation results, obstacle distance measurement results, obstacle speed measurement results, and obstacle length, width, height and other accuracy evaluation results.

3. The method for truth value evaluation based on full FOV annotation according to claim 2, characterized in that: In step S402, the 2D box of the true value result DATA1 is converted to the reference coordinate system of the re-injected image through the virtual internal parameter matrix and the new internal parameter matrix.

4. A device for truth value evaluation based on full FOV annotation, characterized in that: The modules include: A project data acquisition module, used to determine a project, acquire corresponding camera group video data according to the project, and generate camera group picture data DATA according to the camera group video data; A true value result generating module, used for generating a true value result DATA1 according to the camera group picture data DATA; A refill result generating module, used for generating a refill result DATA2 according to the camera group picture data DATA; A true value evaluation module, used for calculating an evaluation result according to the true value result DATA1 and the re-injection result DATA2; Wherein, in the project data acquisition module, the data source of the camera group video data includes a combination of one or more camera data of multiple front-view cameras, rear-view cameras, periscopic cameras, surround-view cameras, monocular cameras, and binocular cameras; The true value result generation module specifically includes the following submodules: The first submodule of the true value result generation is used to obtain the pixel width W of a viewing angle of any camera group image data DATA and the camera horizontal viewing angle FOV, and calculate the new focal length new_fx according to the pixel width W and the camera horizontal viewing angle FOV; The true value result generates a second submodule, which is used to replace the new focal length new_fx into the original camera intrinsic parameter matrix to form a new intrinsic parameter matrix; The third submodule of the true value result generation is used to convert the original image into a dedistorted full FOV image through the original internal parameter matrix, the new internal parameter matrix and the distortion coefficient; A fourth submodule for generating a true value result is used to repeatedly execute the first submodule for generating a true value result, the second submodule for generating a true value result, and the third submodule for generating a true value result, to obtain a plurality of full FOV images of different viewing angles and to mark 2D information of obstacles on the full FOV images of different viewing angles; The fifth submodule of the true value result generation is used to obtain the laser point cloud through the laser radar and mark the 3D information on the laser point cloud; A sixth submodule for generating a true value result is used to fuse the 2D information and the 3D information to generate a true value result DATA1; The 2D information includes box information and obstacle information, and the 3D information includes distance and length, width and height; The recharging result generation module specifically includes the following submodules: The first submodule for generating the feedback result is used to perform a distortion removal process on the camera group image data DATA to generate the distortion-removed data data-1; The second submodule of the refill result generation is used to crop and scale the data data-1 to generate data data-2; The third submodule for generating the reinjection result is used to generate the reinjection result DATA2 by passing the data data-2 through the model and post-processing; In the first submodule of the feedback result generation, the camera group image data DATA is subjected to distortion removal processing through the original camera internal parameter matrix, the distortion coefficient and the virtual internal parameter matrix set by the project to form the distortion-removed data data-1.

5. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for truth value evaluation based on a full FOV annotation method as described in any one of claims 1 to 3 is implemented.

6. An electronic device, characterized in that: It includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements a method for truth value evaluation based on a full FOV annotation method as described in any one of claims 1 to 3.

Images