Vehicle key point marking method and device, equipment, storage medium and program product

Abstract

The invention provides a vehicle key point marking method and device, equipment, a storage medium and a program product, and relates to the technical field of data processing. According to the vehicle key point labeling method, a target vehicle detection frame is determined based on a first image tensor by adopting a target detection model, and a target view angle category and a target vehicle bounding box corresponding to a target vehicle are further determined based on the target vehicle detection frame; vehicle key point labeling deviation caused by visual angle category misjudgment or inaccurate target vehicle bounding box is avoided, the vehicle key point labeling accuracy is improved, and the consistency of vehicle key point labeling results is ensured.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method, apparatus, device, storage medium, and program product for marking key points on a vehicle. Background Technology

[0002] In autonomous driving technology, the annotation of key vehicle points (such as wheelbase, body width, chassis contact points, etc.) is the foundation for building 3D vehicle models and realizing vehicle attitude estimation.

[0003] In related technologies, a single model is typically used for automated annotation of vehicle key points. Specifically, the vehicle image is input into an automated annotation model to automatically annotate the vehicle key points, resulting in the vehicle key points output by the automated annotation model.

[0004] However, the method of automatically labeling key points of vehicles using the above single model has the problem of low labeling accuracy. Summary of the Invention

[0005] This application provides a method, apparatus, device, storage medium, and program product for labeling key points of vehicles, in order to solve the problem of low labeling accuracy in the related art of automatically labeling key points of vehicles using a single model.

[0006] In a first aspect, this application provides a method for annotating vehicle key points, comprising: acquiring a vehicle image corresponding to a target vehicle, and normalizing the vehicle image to obtain a first image tensor corresponding to the vehicle image; inputting the first image tensor into a target detection model for vehicle detection to obtain at least one vehicle detection box output by the target detection model, and determining a target vehicle detection box within the at least one vehicle detection box; determining the target view category corresponding to the target vehicle based on the target vehicle detection box; determining the target vehicle bounding box corresponding to the target vehicle detection box based on the target vehicle detection box; and annotating the vehicle key points corresponding to the target vehicle in the first image tensor based on the target view category and the target vehicle bounding box.

[0007] In one possible implementation, determining the target view category corresponding to the target vehicle based on the target vehicle detection box includes: cropping a first image tensor based on the target vehicle detection box to obtain a second image tensor corresponding to the target vehicle detection box; inputting the second image tensor into a view classification model to perform vehicle view category recognition to obtain a view category probability vector output by the view classification model; and determining the target view category based on the view category probability vector.

[0008] In one possible implementation, determining the target view category based on the view category probability vector includes: determining whether the first view category corresponding to the first view category probability is a preset view category, wherein the first view category probability is the view category probability with the largest probability value in the view category probability vector; if the first view category is a preset view category, then determining whether the first view category probability is greater than or equal to a preset probability threshold; if the first view category probability is greater than or equal to the preset probability threshold, then determining the first view category as the target view category; if the first view category probability is less than the preset probability threshold, then determining the second view category corresponding to the second view category probability as the target view category, wherein the second view category probability is less than the first view category probability, and the second view category is a preset view category.

[0009] In one possible implementation, the vehicle key point annotation method further includes: if the first view category is not a preset view category, or the probability of the first view category is less than a preset probability threshold, then a rollback flag is recorded.

[0010] In one possible implementation, determining the target vehicle bounding box corresponding to the target vehicle detection box based on the target vehicle detection box includes: performing a linear transformation on the target vehicle detection box to obtain a cue vector corresponding to the target vehicle detection box; inputting the cue vector into an image segmentation model to generate a vehicle binary mask corresponding to the target vehicle detection box; performing morphological processing on the vehicle binary mask to obtain the target vehicle binary mask; and extracting the orthogonal bounding rectangle corresponding to the vehicle pixels based on the target vehicle binary mask, and determining the orthogonal bounding rectangle as the target vehicle bounding box.

[0011] In one possible implementation, the vehicle key points corresponding to the target vehicle in the first image tensor are labeled according to the target view category and the target vehicle bounding box, including: labeling the vehicle key points corresponding to the target vehicle in the first image tensor according to the geometric rules corresponding to the target view category and the target vehicle bounding box.

[0012] In one possible implementation, after annotating the vehicle key points corresponding to the target vehicle in the first image tensor, the vehicle key point annotation method further includes: calculating the vehicle pixel size corresponding to the vehicle key points based on the annotated vehicle key points; and mapping the vehicle pixel size to the vehicle physical size according to a preset scaling factor, wherein the preset scaling factor is dynamically adjusted based on calibration data.

[0013] Secondly, this application provides a vehicle key point marking device, comprising:

[0014] The acquisition module is used to acquire the vehicle image corresponding to the target vehicle and perform normalization processing on the vehicle image to obtain the first image tensor corresponding to the vehicle image.

[0015] The detection module is used to input the first image tensor into the target detection model to detect vehicles, obtain at least one vehicle detection box output by the target detection model, and determine the target vehicle detection box in the at least one vehicle detection box;

[0016] The first determination module is used to determine the target view category corresponding to the target vehicle based on the target vehicle detection box;

[0017] The second determining module is used to determine the target vehicle bounding box corresponding to the target vehicle detection box based on the target vehicle detection box.

[0018] The annotation module is used to annotate the vehicle key points corresponding to the target vehicle in the first image tensor according to the target view category and the target vehicle bounding box.

[0019] In one possible implementation, the first determining module is specifically used to: crop a first image tensor based on the target vehicle detection box to obtain a second image tensor corresponding to the target vehicle detection box; input the second image tensor into a view classification model to identify the vehicle view category, and obtain the view category probability vector output by the view classification model; and determine the target view category based on the view category probability vector.

[0020] In one possible implementation, the first determining module is further configured to: determine whether the first view category corresponding to the first view category probability is a preset view category, wherein the first view category probability is the view category probability with the largest probability value in the view category probability vector; if the first view category is a preset view category, then determine whether the first view category probability is greater than or equal to a preset probability threshold; if the first view category probability is greater than or equal to the preset probability threshold, then determine that the first view category is the target view category; if the first view category probability is less than the preset probability threshold, then determine that the second view category corresponding to the second view category probability is the target view category, wherein the second view category probability is less than the first view category probability, and the second view category is the preset view category.

[0021] In one possible implementation, the first determining module is further configured to: record a backtracking flag when the first view category is not a preset view category, or when the probability of the first view category is less than a preset probability threshold.

[0022] In one possible implementation, the second determining module is specifically used for: performing a linear transformation on the target vehicle detection box to obtain a cue vector corresponding to the target vehicle detection box; inputting the cue vector into an image segmentation model to generate a vehicle binary mask corresponding to the target vehicle detection box; performing morphological processing on the vehicle binary mask to obtain a target vehicle binary mask; and extracting the orthogonal bounding rectangle corresponding to the vehicle pixels based on the target vehicle binary mask, and determining the orthogonal bounding rectangle as the target vehicle bounding box.

[0023] In one possible implementation, the annotation module is specifically used to: annotate the vehicle key points corresponding to the target vehicle in the first image tensor according to the geometric rules corresponding to the target view category and the target vehicle bounding box.

[0024] In one possible implementation, after annotating the vehicle key points corresponding to the target vehicle in the first image tensor, the vehicle key point annotation device further includes a mapping module (not shown), which is used to: calculate the vehicle pixel size corresponding to the vehicle key points based on the annotated vehicle key points; and map the vehicle pixel size to the vehicle physical size according to a preset scaling factor, wherein the preset scaling factor is dynamically adjusted based on calibration data.

[0025] Thirdly, this application provides an electronic device, including: a processor and a memory communicatively connected to the processor; the memory stores computer-executable instructions; the processor executes the computer-executable instructions stored in the memory to implement the vehicle key point annotation method provided in the first aspect above.

[0026] Fourthly, this application provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the vehicle key point annotation method provided in the first aspect above.

[0027] Fifthly, this application provides a computer program product, comprising: a computer program that, when executed by a processor, implements the vehicle key point annotation method provided in the first aspect above.

[0028] The vehicle keypoint annotation method, apparatus, device, storage medium, and program product provided in this application involve acquiring a vehicle image corresponding to a target vehicle, normalizing the vehicle image to obtain a first image tensor, inputting the first image tensor into a target detection model for vehicle detection, obtaining at least one vehicle detection box output by the target detection model, identifying the target vehicle detection box within the at least one vehicle detection box, determining the target vehicle detection box, determining the target view category corresponding to the target vehicle based on the target vehicle detection box, determining the target vehicle bounding box corresponding to the target vehicle detection box, and finally annotating the vehicle keypoints corresponding to the target vehicle in the first image tensor based on the target view category and the target vehicle bounding box. This application, by employing a target detection model, determining the target vehicle detection box based on the first image tensor, and further determining the target view category and target vehicle bounding box based on the target vehicle detection box, avoids deviations in vehicle keypoint annotation due to misjudgment of view category or inaccurate target vehicle bounding box, improves the accuracy of vehicle keypoint annotation, and ensures the consistency of vehicle keypoint annotation results. Attached Figure Description

[0029] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0030] Figure 1 A flowchart illustrating the vehicle key point annotation method provided in this application embodiment. Figure 1 ;

[0031] Figure 2 A schematic diagram of a vehicle-facing frontal view provided for an embodiment of this application;

[0032] Figure 3 A schematic diagram showing the rearward view of a vehicle as provided in an embodiment of this application;

[0033] Figure 4 A schematic diagram of a vehicle-facing side view provided for an embodiment of this application;

[0034] Figure 5 A flowchart illustrating the vehicle key point annotation method provided in this application embodiment. Figure 2 ;

[0035] Figure 6 A flowchart illustrating the vehicle key point annotation method provided in this application embodiment. Figure 3 ;

[0036] Figure 7 This is a schematic diagram of the structure of the vehicle key point device provided in the embodiments of this application;

[0037] Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0038] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0039] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0040] In related technologies, the annotation of vehicle key points typically relies on manual annotation or automated annotation based on a single model. On the one hand, manual annotation involves annotators manually labeling vehicle key points in vehicle images. However, manual annotation suffers from low efficiency and inconsistencies in key point data due to subjective biases among annotators, resulting in low annotation accuracy. On the other hand, in automated annotation of vehicle key points using a single model, the model is trained on a large amount of vehicle image data with manually annotated key points. Manual annotation of key points is susceptible to subjective influences, making it difficult to guarantee the consistency of key points in the training data. This inconsistency can then be passed on to the model training process. Therefore, the aforementioned methods of automated annotation of vehicle key points using a single model suffer from low annotation accuracy.

[0041] Based on the problems existing in related technologies, the embodiments of this application adopt a target detection model, which determines the target vehicle detection box based on image tensors, and further determines the target view category and target vehicle bounding box corresponding to the target vehicle based on the target vehicle detection box. This avoids the deviation of vehicle key point annotation due to misjudgment of view category or inaccurate target vehicle bounding box, improves the accuracy of vehicle key point annotation, and ensures the consistency of vehicle key point annotation results.

[0042] The application scenarios of the embodiments of this application will be described below first.

[0043] The vehicle key point annotation method provided in this application is applicable to fields such as data preprocessing for autonomous driving systems, intelligent traffic management, and in-vehicle vision-assisted driving. In autonomous driving technology, vehicle key point annotation (such as wheelbase, vehicle width, chassis contact points, etc.) is fundamental for constructing 3D vehicle models and achieving vehicle attitude estimation. For example, during the training process of an autonomous driving system, accurate vehicle key point annotation is required on massive amounts of vehicle images to generate high-quality datasets for training deep learning models; in intelligent traffic management systems, vehicle key point information can be used to monitor traffic flow in real time and identify illegal parking or abnormal driving behavior.

[0044] The specific implementation of the vehicle key point annotation method provided in this application will be described in detail below with reference to specific embodiments.

[0045] Figure 1 A flowchart illustrating the vehicle key point annotation method provided in this application embodiment. Figure 1 .like Figure 1 As shown, the specific implementation of this vehicle key point annotation method may include the following steps:

[0046] S101, acquire the vehicle image corresponding to the target vehicle, and normalize the vehicle image to obtain the first image tensor corresponding to the vehicle image.

[0047] For example, the vehicle image can be an image containing the target vehicle. The pixel value corresponding to each pixel in the vehicle image can be represented as... .

[0048] For example, the vehicle image may also contain other vehicles besides the target vehicle. However, it is understood that the area occupied by the target vehicle in the vehicle image is larger than the area occupied by other vehicles.

[0049] For example, vehicle images can be acquired using sensors such as cameras, lidar, and millimeter-wave radar.

[0050] One possible way to obtain vehicle images corresponding to a target vehicle is for the user to upload vehicle images acquired by sensors such as cameras, lidar, and millimeter-wave radar to a terminal device or server.

[0051] For example, one possible implementation of normalizing a vehicle image is to normalize the pixel values ​​of the vehicle image into a first image tensor of uniform size. For instance, if the vehicle image resolution is 1920×1080, after normalization, it is converted into a first image tensor in the range [0, 1].

[0052] For example, the first image tensor Where H represents the height of the first image tensor and W represents the width of the first image tensor, and the number of channels of the first image tensor is 3 (e.g., RGB).

[0053] Understandably, normalizing vehicle images can eliminate brightness differences between different vehicle images, providing a unified input scale for subsequent vehicle viewpoint category recognition and target vehicle bounding box detection and segmentation.

[0054] S102, input the first image tensor into the target detection model to perform vehicle detection, obtain at least one vehicle detection box output by the target detection model, and determine the target vehicle detection box within the at least one vehicle detection box.

[0055] For example, the object detection model can be a YOLO series network model. This application does not limit the specific type of object detection model; it can be determined according to actual application requirements.

[0056] For example, each vehicle detection frame in at least one vehicle detection frame can be represented as .in, This indicates the center coordinates of the vehicle detection frame. This indicates the width of the vehicle detection frame. This indicates the height of the vehicle detection frame.

[0057] Understandable, These are the normalized coordinates corresponding to the first image tensor. is the normalized scale corresponding to the first image tensor.

[0058] For example, the target vehicle detection box can be the vehicle detection box with the largest area among at least one vehicle detection box.

[0059] For example, the area corresponding to each vehicle detection frame can be represented as: .

[0060] It is understandable that the target vehicle detection box can be a vehicle detection box that contains the target vehicle.

[0061] For example, in a multi-vehicle scenario, at least one other vehicle detection box can be retained as needed to achieve automated annotation of the key points of each vehicle in the multi-vehicle scenario.

[0062] S103, Based on the target vehicle detection box, determine the target view category corresponding to the target vehicle.

[0063] For example, the target view category can be any one of the frontal view, rearal view, or side view facing the target vehicle.

[0064] One possible implementation of this step is to use a view classification model to identify the view category of the target vehicle detection box, and determine the target view category based on the view category identification results and the backoff mechanism.

[0065] Understandably, determining the target view category corresponding to the target vehicle based on the target vehicle detection bounding box can improve the accuracy of the target view category.

[0066] S104, Based on the target vehicle detection box, determine the target vehicle bounding box corresponding to the target vehicle detection box.

[0067] For example, the bounding box of the target vehicle can be the smallest bounding orthogonal rectangle containing all the pixels corresponding to the target vehicle in the first image tensor.

[0068] In this step, one possible implementation is to obtain a pixel-level bounding box that accurately fits the vehicle outline, i.e., the target vehicle bounding box, based on the target vehicle detection box, through pixel-level image segmentation and morphological optimization.

[0069] S105, based on the target view category and the target vehicle bounding box, annotate the vehicle key points corresponding to the target vehicle in the first image tensor.

[0070] For example, the target vehicle bounding box can be represented as:

[0071]

[0072] in, This represents the pixel coordinates corresponding to the bottom left corner of the target vehicle's bounding box. This represents the pixel coordinates corresponding to the upper right corner of the target vehicle's bounding box.

[0073] Figure 2 This is a schematic diagram of a vehicle-facing frontal view provided in an embodiment of this application. When the target view category is a vehicle-facing frontal view, such as... Figure 2 As shown, the corresponding key vehicle points include the vehicle top midpoint K, the left midpoint I and the right midpoint J based on the vehicle height, and the vehicle chassis contact points L and M.

[0074] In this step, when the target view category is a frontal view facing the vehicle, one possible implementation is: based on the coordinates corresponding to the target vehicle's bounding box, calculate as follows: Figure 2 The diagram shows the vehicle's top midpoint K, left midpoint I based on the vehicle's height, and right midpoint J. Then, based on the bottom edge pixels of the target vehicle's bounding box, a pixel traversal is performed upwards to find the vehicle's foreground pixel closest to the bottom edge pixel. The foreground pixel distance d between the vehicle's foreground pixel and the bottom edge pixel is obtained. When the foreground pixel distance d is less than or equal to an empirical threshold, the vehicle's foreground pixel is determined as shown below. Figure 2 The diagram shows the chassis contact points L and M of the vehicle. The foreground pixels are the points in the binary mask of the target vehicle where the pixel value is 1. The foreground pixel distance is an indicator of the vertical pixel distance from the bottom edge pixel of the target vehicle's bounding box to the main body of the target vehicle.

[0075] For example, the empirical threshold can be 8. This application does not limit the size of the empirical threshold; it can be determined based on actual application requirements.

[0076] Understandably, when the distance d between the foreground pixels is greater than the empirical threshold, it indicates that there is no vehicle chassis result at the location corresponding to the foreground pixel of the vehicle, and it needs to be skipped.

[0077] Optionally, the pixel width of the target vehicle can be calculated based on the left midpoint I and the right midpoint J of the vehicle height.

[0078] For example, the pixel width of the vehicle body can be represented as:

[0079]

[0080] in, Indicates the pixel width of the car body. This represents the x-coordinate of the pixel corresponding to the midpoint J on the right. This represents the x-coordinate of the pixel corresponding to the left midpoint I.

[0081] Figure 3 This is a schematic diagram of a vehicle-oriented rearward view provided in an embodiment of this application. When the target view category is a vehicle-oriented rearward view, such as... Figure 3 As shown, the corresponding key vehicle points include the midpoint P at the top of the vehicle, the midpoint N on the left side and the midpoint O on the right side based on the vehicle height, and the chassis contact points Q and R.

[0082] When the target view category is a rear view facing the vehicle, the corresponding vehicle key point annotation method is similar to the implementation method described above when the target view category is a front view facing the vehicle, and will not be repeated here.

[0083] Figure 4 This is a schematic diagram of a vehicle-facing side view provided in an embodiment of this application. When the target view category is a vehicle-facing side view, such as... Figure 4 As shown, the corresponding key points of the vehicle include the center point A of the left wheel, the center point B of the right wheel, the perpendicular point E corresponding to point A, the perpendicular point F corresponding to point B, the vehicle vertex G, the perpendicular point H corresponding to point G, the left side point of the lower edge of the vehicle side window C, and the right side point of the lower edge of the vehicle side window D.

[0084] For example, point C on the left side of the lower edge of the vehicle side window and point D on the right side of the lower edge of the vehicle side window can be the intersection points of the vehicle side window and the bottom of the vehicle body.

[0085] In this step, when the target view category is a frontal side view facing the vehicle, one possible implementation is as follows: First, use the Hough circle detection algorithm to identify the circular contour of the wheel within the region corresponding to the target vehicle's bounding box, and calculate the set of wheel center points. (e.g., center point A of the left wheel and center point B of the right wheel), set the wheel centers together. The center points of the left and right wheels of the vehicle are obtained by sorting the center points of the wheels according to their pixel x-coordinates; secondly, based on the bottom straight line of the target vehicle... Calculate the perpendicular point E from the center point A of the left wheel of the vehicle, and the perpendicular point E corresponding to the center point B of the right wheel of the vehicle. Then, based on the bottom edge pixels of the target vehicle bounding box, perform pixel traversal upwards to find the left side point C and the right side point D of the bottom edge of the vehicle side window. Based on the top edge pixels of the target vehicle bounding box, perform pixel traversal downwards to find the vehicle foreground pixel closest to the top edge pixel, i.e., the vehicle vertex G, and calculate the perpendicular point H of the vehicle vertex G on the bottom edge line of the target vehicle.

[0086] For example, the pixel coordinates of the perpendicular point E can be represented as: , .in, This represents the x-coordinate of the perpendicular point E in pixels. This represents the x-coordinate of point A in pixels. This represents the pixel ordinate of point A. This represents the pixel ordinate of the perpendicular point E. Represents the bottom straight line The corresponding slope,

[0087] Represents the bottom straight line The corresponding intercept. The calculation method for the pixel coordinates of the perpendicular foot point F and the perpendicular foot point H is similar to that of the perpendicular foot point E, and will not be repeated here.

[0088] Optionally, the vehicle body pixel height corresponding to the target vehicle can be calculated based on the vehicle vertex G and the perpendicular point H.

[0089] For example, the vehicle body pixel height can be represented as:

[0090]

[0091] in, Indicates the pixel height of the vehicle body. This represents the pixel ordinate of the perpendicular point H. This represents the pixel ordinate of the vehicle vertex G.

[0092] Optionally, the pixel wheelbase of the target vehicle can be calculated based on the center point A of the left wheel and the center point B of the right wheel.

[0093] For example, pixel wheel spacing can be expressed as:

[0094]

[0095] in, Indicates pixel wheel spacing. This represents the x-coordinate of the pixel corresponding to the center point A of the vehicle's left wheel. This represents the pixel ordinate of the center point A of the vehicle's left wheel. This represents the x-coordinate of the pixel corresponding to the center point B of the vehicle's right wheel. This represents the pixel coordinate of the center point B of the right wheel of the vehicle.

[0096] In this embodiment, a target detection model is employed to determine the target vehicle detection box based on a first image tensor. Furthermore, based on the target vehicle detection box, the target viewpoint category and the target vehicle bounding box are determined. This avoids deviations in vehicle keypoint annotation due to misjudgment of viewpoint category or inaccurate target vehicle bounding boxes, improving the accuracy of vehicle keypoint annotation and ensuring consistency of the annotation results. Simultaneously, it automates the entire process from vehicle image acquisition to vehicle keypoint annotation, reducing manual intervention, lowering the cost and subjective bias of manual annotation, improving annotation efficiency and automation, and reducing labor costs.

[0097] The following is combined with Figure 5 The specific implementation method of step S103, which determines the target view category corresponding to the target vehicle based on the target vehicle detection box, is explained in detail.

[0098] Figure 5 A flowchart illustrating the vehicle key point annotation method provided in this application embodiment. Figure 2 .like Figure 5 As shown, the specific implementation of determining the target view category corresponding to the target vehicle based on the target vehicle detection box in this vehicle key point annotation method may include the following steps:

[0099] S501, based on the target vehicle detection box, crop the first image tensor to obtain the second image tensor corresponding to the target vehicle detection box.

[0100] In this step, one possible implementation is to crop the image tensor in the corresponding region of the first image tensor based on the center normalized coordinates, normalized width, and normalized height of the target vehicle detection box, thereby obtaining the second image tensor corresponding to the target vehicle detection box.

[0101] S502, input the second image tensor into the view classification model to identify the vehicle view category, and obtain the view category probability vector output by the view classification model.

[0102] For example, a view classification model includes a view classifier. And the normalized exponential function softmax. Where, the view classifier... The backbone of the EfficientNet convolutional neural network, based on transfer learning, is adopted.

[0103] In one possible implementation, a view classifier It can be trained by transferring the parameters of the EfficientNet convolutional neural network pre-trained on a large-scale general dataset to a viewpoint classifier. Then, the parameters are fine-tuned and trained for different vehicle perspectives (such as a 45° perspective facing the vehicle, a frontal perspective facing the vehicle, a rearal perspective facing the vehicle, and a side perspective facing the vehicle) to obtain a perspective classifier. .

[0104] One possible implementation of this step is to input the second image tensor into the viewpoint classifier. Perform vehicle view category recognition to obtain a view classifier. The output view category score vector is then normalized using the normalized exponential function softmax to obtain the view category probability vector output by the normalized exponential function softmax.

[0105] For example, the view category probability vector can be represented as:

[0106]

[0107] in, Represents the second image tensor. This represents the probability vector of view category.

[0108] S503, determine the target view category based on the view category probability vector.

[0109] For example, the target view category can be any one of the following: a frontal view of the vehicle, a rearward view of the vehicle, or a side view of the vehicle.

[0110] Alternatively, one possible implementation of this step is to determine the first-view category probability. Corresponding first-person perspective category Whether it is a preset view category, the probability of the first view category is the view category probability with the highest probability value in the view category probability vector; if the first view category is a preset view category, then determine whether the probability of the first view category is greater than or equal to a preset probability threshold. If the probability of the first-person perspective category is greater than or equal to the preset probability threshold ( If the probability of the first-view category is less than the preset probability threshold, then the first-view category is determined as the target view category; if the probability of the first-view category is less than the preset probability threshold, then the first-view category is determined as the target view category. Then determine the probability of the second-view category. Corresponding second-person perspective category The target view category is defined as follows: the probability of the second view category is less than that of the first view category, and the second view category is defined as the preset view category.

[0111] For example, the preset view categories can be a front view facing the vehicle, a rear view facing the vehicle, and a side view facing the vehicle.

[0112] It should be noted that the vehicle key point annotation method provided in this application mainly targets three view categories for vehicle key point annotation: the front view of the vehicle, the rear view of the vehicle, and the side view of the vehicle.

[0113] Understandably, the probability of each view category in the view category probability vector can be used to represent the confidence level of that view category. Correspondingly, the preset probability threshold can also be described as a confidence threshold.

[0114] The embodiments of this application do not limit the size of the preset probability threshold; it can be determined according to the actual application requirements.

[0115] Understandably, the first-person perspective category is the preferred category, and the second-person perspective category is the second-best category.

[0116] In this embodiment, a second image tensor corresponding to the target vehicle detection box is obtained by cropping a first image tensor based on the target vehicle detection box, and the target view category is determined based on the second image tensor, thereby improving the accuracy of the target view category. Furthermore, the second image tensor is input into a view classification model to identify the vehicle view category, thereby obtaining the view category probability vector output by the view classification model. The target view category is then determined based on the view category probability vector, thereby reducing the misclassification rate of the target view category and improving the accuracy of the target view category.

[0117] Optionally, the vehicle key point annotation method provided in this application embodiment further includes: recording a rollback flag if the first view category is not a preset view category, or the probability of the first view category is less than a preset probability threshold.

[0118] In some embodiments, if the first view category is not a preset view category, a fallback logic of "falling back from the preferred category to the second-best category" is triggered, and a fallback flag is recorded to indicate that the view classification for the second image tensor is not "direct output with high confidence" and a downgrade decision has been made. In some embodiments, if the first view category is a preset view category and the probability of the first view category is less than a preset probability threshold, a fallback logic of "falling back from the preferred category to the second-best category" is triggered, and a fallback flag is recorded to indicate that the view classification for the second image tensor is not "direct output with high confidence" and a downgrade decision has been made.

[0119] Understandably, the purpose of recording the rollback flag is to mark the state in which the system triggers the rollback strategy during the view classification process, so that the view category corresponding to the target vehicle in the second image tensor can be specifically verified and optimized in subsequent stages to ensure the reliability of view category judgment.

[0120] Compared to related technologies that use a single model for automated annotation of vehicle key points, key point drift is prone to occur under complex lighting conditions such as day-night transitions, backlighting, and strong reflections, leading to misclassification of viewpoint categories. This application's embodiment employs a confidence-backoff-driven viewpoint category stabilization mechanism. Based on the viewpoint classification model, when the probability of the first viewpoint category (preferred category) is lower than a preset probability threshold, it automatically backoffs to the second viewpoint category (suboptimal category) and records the backoff flag. This ensures that the correct geometric rules corresponding to the target viewpoint category can be selected in vehicle images corresponding to the three viewpoint categories: front view, rear view, and side view, suppressing the chain reaction of misclassifications.

[0121] The following is combined with Figure 6 The specific implementation of step S204, which determines the target vehicle bounding box corresponding to the target vehicle detection box, is explained in detail.

[0122] Figure 6 A flowchart illustrating the vehicle key point annotation method provided in this application embodiment. Figure 3 The specific implementation of this vehicle key point annotation method, which determines the target vehicle bounding box corresponding to the target vehicle detection box based on the target vehicle detection box, may include the following steps:

[0123] S601, perform linear transformation on the target vehicle detection box to obtain the prompt vector corresponding to the target vehicle detection box.

[0124] In this step, one possible implementation is: [Detecting the bounding box corresponding to the target vehicle]. A linear transformation is performed to obtain a pixel box of the same size as the target detection box, and the cue vector is generated based on the pixel coordinates corresponding to the pixel box.

[0125] For example, the pixel coordinates corresponding to the pixel box It can be represented as:

[0126]

[0127]

[0128] For example, ( ) can be the pixel coordinates corresponding to the bottom left corner of the pixel box, ( () can be the pixel coordinates corresponding to the upper right corner of the pixel box.

[0129] For example, the area corresponding to a pixel box can be represented as:

[0130]

[0131] For example, a cue vector can be represented as .

[0132] S602, input the prompt vector into the image segmentation model to generate a vehicle binary mask corresponding to the target vehicle detection box.

[0133] For example, the image segmentation model can be the Segment Anything model (SAM).

[0134] In one possible implementation, the cue vector is input into the SAM to obtain the vehicle binary mask output by the SAM. .

[0135] For example, in a vehicle binary mask, the number 1 represents the pixel of the target vehicle, and the number 0 represents the background pixel.

[0136] S603 performs morphological processing on the vehicle binary mask to obtain the target vehicle binary mask.

[0137] For example, morphological processing can be an erosion-dilation operation.

[0138] In one possible implementation, an erosion-dilation operation is performed on the vehicle binary mask to remove isolated noise points in the vehicle binary mask, ensuring the continuity of the vehicle contour corresponding to the target vehicle, and thus obtaining the target vehicle binary mask.

[0139] For example, the binary mask for the target vehicle can be represented by the following formula:

[0140]

[0141] in, This represents the binary mask for the target vehicle. The symbol represents the erosion operation to eliminate isolated noise. Representing morphological structural elements, This indicates an expansion operation symbol used to restore the continuity of the vehicle's outline.

[0142] S604. Based on the binary mask of the target vehicle, extract the orthogonal bounding rectangle corresponding to the vehicle pixels, and determine the orthogonal bounding rectangle as the bounding box of the target vehicle.

[0143] In this step, one possible implementation is to find the extreme value of the set of vehicle foreground pixels (non-zero pixel set) in the binary mask of the target vehicle to obtain the orthogonal bounding rectangle corresponding to the vehicle pixels.

[0144] For example, the set of foreground pixels of a vehicle can be represented as:

[0145]

[0146] For example, an orthogonal circumscribed rectangle can be represented as:

[0147]

[0148] in,( () represents the coordinates of the lower left corner of the orthogonal circumscribed rectangle. () represents the coordinates of the upper right corner of the orthogonal circumscribed rectangle.

[0149] It should be noted that when the target view category is a frontal side view facing the vehicle, only the following can be used. The left and right boundaries of the region are updated to reduce the influence of the roof attachments, and the equations of the vertical lines in the upper, lower, left and right directions are constructed for the above perpendicular calculation.

[0150] In this embodiment, a cue vector corresponding to the target vehicle detection box is obtained by performing a linear transformation on the target vehicle detection box. The cue vector is then input into an image segmentation model to generate a binary mask for the vehicle corresponding to the target vehicle detection box. The binary mask is further subjected to morphological processing to obtain a binary mask for the target vehicle. Then, based on the binary mask for the target vehicle, the orthogonal bounding rectangle corresponding to the vehicle pixels is extracted, and the orthogonal bounding rectangle is determined as the target vehicle bounding box, thereby improving the accuracy of the target vehicle bounding box. Finally, vehicle key points are labeled based on the target vehicle bounding box, thereby improving the labeling accuracy.

[0151] Compared to related technologies that use a single model for automated vehicle keypoint annotation, which suffer from inaccurate viewpoint category determination due to factors such as partial vehicle occlusion and high similarity between the front and rear of the vehicle, thus affecting the selection of geometric rules corresponding to the viewpoint category and resulting in low annotation accuracy, this application employs a collaborative mechanism of object detection and masking to mitigate the effects of lighting and occlusion. Furthermore, addressing the challenge of vehicle keypoint drift under strong reflections and shadows, an object detection model provides coarse localization of the vehicle contour corresponding to the target vehicle, and an image segmentation model outputs a high-quality binary vehicle mask. The combination of these two methods forms a robust contour, providing stable input for subsequent geometric inference of vehicle keypoint annotation, maintaining accuracy even in the face of partial occlusion.

[0152] Optionally, one possible implementation of step S105, which involves labeling the vehicle key points corresponding to the target vehicle in the image tensor based on the target view category and the target vehicle bounding box, is as follows: label the vehicle key points corresponding to the target vehicle in the first image tensor based on the geometric rules corresponding to the target view category and the target vehicle bounding box.

[0153] For example, when the target view category is a side view facing the vehicle, the corresponding geometric rule can be the Hough circle detection algorithm and perpendicularity calculation; when the target view category is a front view facing the vehicle or a rear view facing the vehicle, the corresponding geometric rule can be a straight line equation.

[0154] The specific implementation of this embodiment is similar to that described above, and will not be repeated here.

[0155] In this embodiment, by annotating the vehicle key points corresponding to the target vehicle in the first image tensor according to the geometric rules corresponding to the target view category and the target vehicle bounding box, the uncertainty caused by occlusion can be reduced by analytical geometric solution. Faced with the difficulty in locating vehicle key points caused by partial obstruction of the outline or high similarity between the front and rear of the vehicle, analytical geometric methods such as Hough circle detection algorithm, straight line equation and perpendicular foot calculation are used to derive vehicle key points such as wheel track, upper and lower boundaries and chassis contact points, ensuring that the vehicle key points are interpretable and verifiable.

[0156] Optionally, the vehicle key point annotation method provided in this application embodiment further includes, after annotating the vehicle key points corresponding to the target vehicle in the first image tensor: calculating the vehicle pixel size corresponding to the vehicle key points based on the annotated vehicle key points; and mapping the vehicle pixel size to the vehicle physical size according to a preset scaling factor, wherein the preset scaling factor is dynamically adjusted based on calibration data.

[0157] For example, the vehicle body pixel size can be the vehicle body pixel width, vehicle body pixel height, and wheelbase pixel dimensions. The specific calculation method is similar to that described above and will not be repeated here.

[0158] For example, the calibration data can be the camera's intrinsic and extrinsic parameters used to capture vehicle images, or the corresponding calibration data of external LiDAR, millimeter-wave radar, etc. used to acquire vehicle images.

[0159] The method for mapping vehicle pixel dimensions to vehicle physical dimensions based on a preset scaling factor can be expressed by the following formula:

[0160]

[0161] in, Indicates the vehicle's physical dimensions. Indicates the vehicle pixel size. This indicates the preset scaling factor.

[0162] Optionally, after mapping the vehicle pixel size to the vehicle physical size according to a preset scaling factor, the vehicle key point annotation method provided in this application embodiment further includes: constructing a projection matrix by combining parameters such as pitch angle, focal length, and imaging principal point to correct the vehicle's three-dimensional geometry.

[0163] Optionally, when calibration data is missing, the vehicle pixel size, vehicle key points, and reliable range are output for external system fusion positioning module or manual refinement, thereby forming a closed-loop adaptive size management mechanism.

[0164] Compared to related technologies that use a single model for automated annotation of vehicle key points, the lack of a reliable size conversion mechanism leads to the accumulation of errors in important parameters such as wheelbase and vehicle width. This application's embodiments mitigate this accumulation by introducing a multi-source scale fusion size conversion framework. Specifically, by introducing a multi-source fusion framework that incorporates preset scaling factors, calibration data, and attitude parameters, the vehicle body pixel dimensions are calibrated from first-order to sub-pixel level. In the absence of calibration, a reliable range can be output for external closed-loop correction, thereby suppressing the accumulation of size errors.

[0165] Optionally, the vehicle key point method provided in this application embodiment further includes: uniformly encapsulating the view category, vehicle key point coordinates, vehicle binary mask and vehicle body pixel size data into a structured record, whose corresponding fields may include view_type, points, bbox, mask, dimensions, etc., for subsequent quality inspection, modeling or document generation, and can generate a visualization map with vehicle key point coverage.

[0166] It should be noted that the vehicle key point annotation method provided in this application embodiment can be implemented through a cloud server. In order to cope with different computing power conditions or business needs, it can also be implemented through edge devices, batch servers or cloud environments, etc., based on lightweight object detection models, image segmentation models, learning-based vehicle key point regression, etc.

[0167] In summary, the vehicle key point annotation method provided in this application has the following beneficial effects:

[0168] 1) Improve annotation accuracy through multi-stage fusion: Through target detection model, image segmentation model and set normalization, high-quality vehicle outline, i.e. target vehicle bounding box, is obtained. Then, the vehicle key points are robustly solved by analytical geometry, so that the size errors of pixel wheel track, body pixel width and height are continuously maintained at the pixel level, avoiding scale drift caused by manual thresholding.

[0169] 2) Strong robustness in complex scenarios: By adopting a view classifier backoff mechanism, it is ensured that the correct template can be triggered for the three types of vehicle images: the front view, the rear view, and the side view. Combined with the image segmentation mask, it has the ability to adapt to strong light, shadow, and occlusion, which significantly reduces the misidentification rate of view categories.

[0170] 3) Improved annotation efficiency and consistency: The fully automated output of vehicle key points, vehicle binary masks and vehicle pixel size parameters reduces repeated manual verification. At the same time, unified geometric solution constraints ensure the consistency of vehicle key point output between different vehicle images, which can be directly connected to quality inspection and modeling systems.

[0171] 4) Modular expansion and convenient replacement: View category classification, object detection, image segmentation and vehicle key point parsing are all encapsulated with interfaces, which can be seamlessly replaced with lightweight or more advanced models, compatible with GPU and CPU environments, and support scene expansion from single vehicle to multi-vehicle.

[0172] 5) Data asset accumulation capability: Output standardized structured records and visualizations to provide high-quality data assets for subsequent training of enhanced models or traceability audits, thereby reducing operation and maintenance costs.

[0173] The following are embodiments of the apparatus described in this application, which can be used to execute the embodiments of the method described in this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the method described in this application.

[0174] Figure 7 This is a schematic diagram of the structure of the vehicle key point device provided in an embodiment of this application. Figure 7 As shown, the vehicle key point device 70 includes an acquisition module 710, a detection module 720, a first determination module 730, a second determination module 740, and a labeling module 750.

[0175] The acquisition module 710 is used to acquire the vehicle image corresponding to the target vehicle and normalize the vehicle image to obtain the first image tensor corresponding to the vehicle image.

[0176] The detection module 720 is used to input the first image tensor into the target detection model to detect vehicles, obtain at least one vehicle detection box output by the target detection model, and determine the target vehicle detection box in the at least one vehicle detection box;

[0177] The first determining module 730 is used to determine the target view category corresponding to the target vehicle based on the target vehicle detection box;

[0178] The second determining module 740 is used to determine the target vehicle bounding box corresponding to the target vehicle detection box based on the target vehicle detection box.

[0179] The annotation module 750 is used to annotate the vehicle key points corresponding to the target vehicle in the first image tensor according to the target view category and the target vehicle bounding box.

[0180] In one possible implementation, the first determining module 730 is specifically used to: crop the first image tensor based on the target vehicle detection box to obtain the second image tensor corresponding to the target vehicle detection box; input the second image tensor into the view classification model to perform vehicle view category recognition to obtain the view category probability vector output by the view classification model; and determine the target view category based on the view category probability vector.

[0181] In one possible implementation, the first determining module 730 is further configured to: determine whether the first view category corresponding to the first view category probability is a preset view category, wherein the first view category probability is the view category probability with the largest probability value in the view category probability vector; if the first view category is a preset view category, then determine whether the first view category probability is greater than or equal to a preset probability threshold; if the first view category probability is greater than or equal to the preset probability threshold, then determine that the first view category is a target view category; if the first view category probability is less than the preset probability threshold, then determine that the second view category corresponding to the second view category probability is a target view category, wherein the second view category probability is less than the first view category probability, and the second view category is a preset view category.

[0182] In one possible implementation, the first determining module 730 is further configured to: record a backtracking flag when the first view category is not a preset view category, or when the probability of the first view category is less than a preset probability threshold.

[0183] In one possible implementation, the second determining module 740 is specifically used for: performing linear transformation on the target vehicle detection box to obtain the cue vector corresponding to the target vehicle detection box; inputting the cue vector into the image segmentation model to generate a vehicle binary mask corresponding to the target vehicle detection box; performing morphological processing on the vehicle binary mask to obtain the target vehicle binary mask; and extracting the orthogonal bounding rectangle corresponding to the vehicle pixels based on the target vehicle binary mask, and determining the orthogonal bounding rectangle as the target vehicle bounding box.

[0184] In one possible implementation, the annotation module 750 is specifically used to: annotate the vehicle key points corresponding to the target vehicle in the first image tensor according to the geometric rules corresponding to the target view category and the target vehicle bounding box.

[0185] In one possible implementation, after annotating the vehicle key points corresponding to the target vehicle in the first image tensor, the vehicle key point annotation device further includes a mapping module (not shown), which is used to: calculate the vehicle pixel size corresponding to the vehicle key points based on the annotated vehicle key points; and map the vehicle pixel size to the vehicle physical size according to a preset scaling factor, wherein the preset scaling factor is dynamically adjusted based on calibration data.

[0186] The vehicle key point marking device provided in this embodiment can be used to execute the method steps of the above method embodiment. The specific implementation and technical effects are similar, and will not be repeated here.

[0187] Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 8 As shown, the electronic device 80 provided in this embodiment includes at least one processor 801 and a memory 802. Optionally, the electronic device 80 further includes a communication component 803. The processor 801, memory 802, and communication component 803 are connected via a bus 804.

[0188] In a specific implementation, at least one processor 801 executes computer execution instructions stored in memory 802, causing at least one processor 801 to perform the above-described method.

[0189] The specific implementation process of processor 801 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0190] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0191] The memory may include random access memory (RAM) and non-volatile memory (NVM), such as at least one disk storage device.

[0192] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0193] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0194] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0195] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0196] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0197] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0198] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0199] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0200] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0201] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0202] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A method for marking key points on a vehicle, characterized in that, include: Obtain the vehicle image corresponding to the target vehicle, and normalize the vehicle image to obtain the first image tensor corresponding to the vehicle image. The first image tensor is input into the target detection model to detect vehicles, and at least one vehicle detection box is obtained from the target detection model. The target vehicle detection box is then determined within the at least one vehicle detection box. Based on the target vehicle detection box, determine the target view category corresponding to the target vehicle; Based on the target vehicle detection box, determine the target vehicle bounding box corresponding to the target vehicle detection box; Based on the target viewpoint category and the target vehicle bounding box, label the vehicle key points corresponding to the target vehicle in the first image tensor.

2. The vehicle key point annotation method according to claim 1, characterized in that, The step of determining the target view category corresponding to the target vehicle based on the target vehicle detection box includes: Based on the target vehicle detection box, the first image tensor is cropped to obtain the second image tensor corresponding to the target vehicle detection box; The second image tensor is input into the view classification model to identify the vehicle view category, and the view category probability vector output by the view classification model is obtained. The target view category is determined based on the view category probability vector.

3. The vehicle key point annotation method according to claim 2, characterized in that, Determining the target view category based on the view category probability vector includes: Determine whether the first view category corresponding to the first view category probability is a preset view category, wherein the first view category probability is the view category probability with the largest probability value in the view category probability vector; If the first view category is the preset view category, then determine whether the probability of the first view category is greater than or equal to the preset probability threshold; If the probability of the first view category is greater than or equal to the preset probability threshold, then the first view category is determined to be the target view category; If the probability of the first view category is less than the preset probability threshold, then the second view category corresponding to the probability of the second view category is determined to be the target view category. If the probability of the second view category is less than the probability of the first view category, then the second view category is the preset view category.

4. The vehicle key point annotation method according to claim 3, characterized in that, Also includes: If the first view category is not the preset view category, or the probability of the first view category is less than the preset probability threshold, then a rollback flag is recorded.

5. The vehicle key point annotation method according to claim 1, characterized in that, The step of determining the target vehicle bounding box corresponding to the target vehicle detection box based on the target vehicle detection box includes: The target vehicle detection box is linearly transformed to obtain the prompt vector corresponding to the target vehicle detection box. The cue vector is input into the image segmentation model to generate a vehicle binary mask corresponding to the target vehicle detection box; Morphological processing is performed on the vehicle binary mask to obtain the target vehicle binary mask; Based on the binary mask of the target vehicle, extract the orthogonal bounding rectangle corresponding to the vehicle pixels, and determine the orthogonal bounding rectangle as the bounding box of the target vehicle.

6. The method for marking key points of a vehicle according to any one of claims 1 to 5, characterized in that, The step of annotating the vehicle key points corresponding to the target vehicle in the first image tensor according to the target view category and the target vehicle bounding box includes: Based on the geometric rules corresponding to the target viewpoint category and the target vehicle bounding box, the vehicle key points corresponding to the target vehicle in the first image tensor are labeled.

7. The method for marking key points of a vehicle according to any one of claims 1 to 5, characterized in that, After annotating the vehicle key points corresponding to the target vehicle in the first image tensor, the method further includes: Calculate the vehicle pixel size corresponding to the vehicle key points obtained from the annotation; The vehicle pixel size is mapped to the vehicle physical size according to a preset scaling factor, which is dynamically adjusted based on calibration data.

8. A vehicle key point marking device, characterized in that, include: The acquisition module is used to acquire the vehicle image corresponding to the target vehicle and perform normalization processing on the vehicle image to obtain the first image tensor corresponding to the vehicle image. The detection module is used to input the first image tensor into the target detection model to detect vehicles, obtain at least one vehicle detection box output by the target detection model, and determine the target vehicle detection box in the at least one vehicle detection box; The first determining module is used to determine the target view category corresponding to the target vehicle based on the target vehicle detection box; The second determining module is used to determine the target vehicle bounding box corresponding to the target vehicle detection box based on the target vehicle detection box; The annotation module is used to annotate the vehicle key points corresponding to the target vehicle in the first image tensor according to the target view category and the target vehicle bounding box.

9. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the vehicle key point annotation method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the vehicle key point annotation method as described in any one of claims 1 to 7.

11. A computer program product, characterized in that, include: A computer program, which, when executed by a processor, implements the vehicle key point annotation method as described in any one of claims 1 to 7.

Images