A key point target detection method and system based on multiple correspondences

Abstract

This invention belongs to the field of object detection and provides a keypoint object detection method and system based on multiple correspondences. Addressing the problem that keypoint-based object detection algorithms cannot effectively distinguish multiple objects when their keypoints overlap, this invention utilizes a multiple correspondence relationship in the network, where one location corresponds to multiple predicted boxes to achieve our goal. To achieve this, the method mainly includes the following two parts. First, the keypoint pooling method is improved. Existing keypoint pooling methods focus on emphasizing keypoints, but this invention considers that keypoint detection should emphasize keypoints while suppressing the surrounding environment. Second, to better match corners to the correct center point, a pointing offset is used to match keypoints. Specifically, corner points are matched to the center point through their offset. To better predict the pointing offset, a corner-to-center module is designed to improve the matching effect.

Description

Technical Field

[0001] This invention belongs to the field of target detection technology, and in particular relates to a key point target detection method and system based on multiple correspondence relationships. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] With the development of deep neural networks, anchor points play a crucial role in improving accuracy and recall in both single-level and two-level object detection algorithms. However, processing a large number of anchor points results in a significant computational burden.

[0004] CornerNet abandons the use of anchor points to predict object positions, instead using the object's two diagonal points as keypoints to label the object, achieving state-of-the-art detection performance in one-stage object detection algorithms. The CenterNet algorithm proposed by Zhou et al. also achieves good detection results by abandoning semantically insufficient corner points and using the center as the keypoint to determine the bounding box position. However, this method has the following drawbacks:

[0005] 1. In a network, keypoints and bounding boxes have a one-to-one relationship, meaning one location corresponds to only one predicted box. When keypoints of multiple targets overlap, keypoint-based target detection algorithms cannot effectively distinguish them. To address the center overlap problem, existing anchor-based target detectors use anchor boxes of different scales for differentiation. However, a fixed set of anchor boxes cannot meet the needs of all datasets, so different datasets require different types of anchor boxes. The drawback of this approach is that anchor box processing requires a large amount of memory and computation. Another solution is to predict multiple keypoints and use a set of keypoints to represent an object. The drawback of this approach is the increased cost of predicting keypoints, while also facing the cost of keypoint matching.

[0006] 2. The matching accuracy between key points directly affects the output of the network's predicted bounding box, which is crucial to the network's prediction performance. Existing methods use embedding vectors to match two key points, but this matching method may cause two similar and close objects to generate incorrect predicted bounding boxes, increasing the false detection rate.

[0007] 3. Existing keypoint pooling methods have been proven effective, but they focus on keypoints without considering the surrounding environment, resulting in low prediction accuracy. Summary of the Invention

[0008] To address at least one of the technical problems in the background art described above, the present invention provides a key point target detection method and system based on multiple correspondence relationships. This method uses center points and corner points with pointing offsets to mark objects through multiple correspondence relationships between key points, so that multiple corner points can match the same center point, and multiple center points can also match the same corner point.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] The first aspect of this invention provides a key point target detection method based on multiple correspondences, comprising the following steps:

[0011] Acquire the image data to be detected;

[0012] Detection is performed based on the image data to be detected and the trained keypoint target detection model to obtain the target detection box;

[0013] The construction process of the key point target detection model includes: extracting center points and corner points using camel pooling; using the multiple correspondence between center points and corner points, marking the target to be detected using center points and corner points with pointing offsets; wherein the multiple correspondence is: multiple corner points match the same center point or multiple center points match the same corner point.

[0014] A second aspect of the present invention provides a target detection system based on key point multiple correspondences, comprising:

[0015] The data acquisition module is used to acquire the image data to be detected.

[0016] The object detection module is used to detect objects based on the image data to be detected and the trained keypoint object detection model, and obtain object detection boxes.

[0017] The construction process of the key point target detection model includes: extracting center points and corner points using camel pooling; using the multiple correspondence between center points and corner points, marking the target to be detected using center points and corner points with pointing offsets; wherein the multiple correspondence is: multiple corner points match the same center point or multiple center points match the same corner point.

[0018] A third aspect of the present invention provides a computer-readable storage medium.

[0019] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a key point target detection method based on multiple correspondences as described above.

[0020] A fourth aspect of the present invention provides a computer device.

[0021] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the key point target detection method based on multiple correspondences as described above.

[0022] Compared with the prior art, the beneficial effects of the present invention are:

[0023] 1. To address the issue of one-to-one relationships in networks, where one location corresponds to only one prediction box, this invention utilizes multiple correspondences between keypoints. It labels objects using center points and corner points with directional offsets. This labeling method allows multiple corner points to match the same center point, and vice versa. By matching corners with centers, multiple correspondences are achieved, mitigating the problem of keypoint overlap.

[0024] 2. Starting from the relationship between the center point and the corner points, model the two key points and use the position information of the two key points to assist in the prediction of the pointing offset, so as to achieve a better matching effect.

[0025] 3. Existing keypoint pooling methods focus on emphasizing keypoints. This invention improves the existing keypoint pooling by designing a camelCase pooling method that is more suitable for keypoints, which enhances positive keypoints and suppresses negative keypoints around them.

[0026] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0027] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0028] Figure 1 This is a schematic diagram of the target detection model provided in an embodiment of the present invention.

[0029] Figure 2 This is a schematic diagram of the camel hump pooling method provided in an embodiment of the present invention.

[0030] Figure 3 This is the CTC module structure provided in the embodiments of the present invention.

[0031] Figures 4(a)-4(b) The objects marked with center point, corner point and pointing offset are provided in the embodiments of the present invention; the dashed arrows represent the pointing offset components on the X-axis and Y-axis respectively.

[0032] Figures 5(a)-5(d)The corner-center matching provided in this embodiment of the invention follows certain rules; wherein, in Figure 5(a), the confidence of the corner point and the center point to which the corner point points must be greater than a threshold; in Figure 5(b), the corner point must be located to the lower right of the center point; in Figure 5(c), a detection box is generated based on the last remaining corner-center pair; in Figure 5(d), increasing the matching range of the corner point can effectively improve the recall rate of the model.

[0033] Figure 6 This is a schematic diagram of the prediction box output provided in an embodiment of the present invention. Detailed Implementation

[0034] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0035] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0036] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0037] As mentioned in the background section, existing methods utilize inherent feature points on feature maps to predict object keypoints, thus eliminating the need for anchor point calculations. However, when the keypoints of multiple targets overlap, existing keypoint-based object detection algorithms cannot effectively distinguish them. The problem arises because a one-to-one relationship exists within the network, meaning that one location corresponds to only one predicted bounding box.

[0038] This invention primarily addresses the problem caused by center overlap by employing multiple correspondences between key points. Therefore, center points and corner points with directional offsets are used to mark the objects to be detected. This form of annotation allows multiple corner points to match the same center point, and vice versa.

[0039] Multiple prediction boxes can be used to correspond to one location to achieve the purpose of prediction.

[0040] To achieve this goal, the inventive points of this invention mainly include the following two parts:

[0041] First, the keypoint pooling method is improved. Existing keypoint pooling methods focus on emphasizing keypoints, while this invention considers not only that keypoint detection should emphasize keypoints but also suppress the surrounding environment. Therefore, the existing keypoint pooling method is improved by designing a camelCase pooling method that is more suitable for keypoints.

[0042] Secondly, in order to better match corner points to the correct center point, a pointing offset is used to match key points. Corner points will match the center point through their offset. In order to better predict the pointing offset, a corner-to-center (CTC) module is designed to improve the matching effect.

[0043] Example 1

[0044] Reference Figure 1 This embodiment provides a target detection method based on multiple correspondences of key points, including the following steps:

[0045] Acquire the image data to be detected;

[0046] Detection is performed based on the image data to be detected and the trained keypoint target detection model to obtain the target detection box;

[0047] The construction process of the key point target detection model includes: extracting center points and corner points using camel pooling; using the multiple correspondence between center points and corner points, marking the target to be detected using center points and corner points with pointing offsets; wherein the multiple correspondence is: multiple corner points match the same center point or multiple center points match the same corner point.

[0048] The overall structure diagram of the model is as follows Figure 1 As shown, the prediction head consists of four parts: Center-Head for predicting the center heatmap, Corner-Head for predicting the corner heatmap, and Offset-Head for fine-tuning the keypoint positions due to rounding operations during downsampling. The predicted position offset is shared by the center point and its matching corner points. Matching-Head is used to predict the pointing offset to match the corner points and the center point.

[0049] Specifically, the steps include:

[0050] Step 1: Acquire the image data to be detected;

[0051] Step 2: Extract center heatmap, corner heatmap, and basic feature map based on the image data to be detected;

[0052] In step 2, the center heatmap and corner heatmap contain rich semantic information about the target to be detected; wherein, the center heatmap contains the center information of the target to be detected, the corner heatmap contains the corner information of the target to be detected, and the basic feature map contains rich feature information about the target to be detected.

[0053] It should be noted that the detected corner point can be any one of the four corner points of the object; in this embodiment, the lower right corner of the object to be tested is selected.

[0054] Step 3: Based on the center heatmap and corner heatmap, extract key points using camel pooling, specifically including:

[0055] like Figure 2 As shown, this embodiment uses central pooling as an example for further explanation.

[0056] Existing keypoint pooling methods, such as center pooling proposed by Duan et al. and corner pooling proposed by Law et al., mainly focus on emphasizing keypoints.

[0057] The present invention takes into account that while emphasizing positive key points, it is also necessary to suppress negative key points around positive key points, because negative key points and positive key points will be strengthened together during the convergence of key points.

[0058] The keypoints are pooled in a certain direction, and the pooling results of multiple directions are added together to obtain the emphasis representation of the keypoints. The result is defined as E.

[0059] Then, combining the idea of ​​spatial attention, max pooling and average pooling are used along the channel axis to integrate key point information from different channels into the same space;

[0060] Based on the results of this integration, negative keypoints are suppressed both horizontally and vertically. In the horizontal direction, a horizontal suppression matrix hm∈Size is defined. H×W ,hm∈Size H×W Fill with 1;

[0061] Find the location of the maximum value in each row;

[0062] Suppose the maximum value v in one of the rows i,r At (i,r),

[0063]

[0064] In the formula, μ is a hyperparameter that determines the degree of inhibition, and γ and σ are inhibition factors.

[0065] According to formula (1), when n∈[i-μ, i+μ], the value v n,r It should be suppressed.

[0066] The actual suppression operation is performed at the corresponding position in the horizontal suppression matrix hm.

[0067] Vertical suppression is similar to horizontal suppression, and we can obtain the vertical suppression matrix vm. Then, we multiply vm and hm at corresponding positions, and define the result as S.

[0068] Multiply E and S to get the output of camelCase pooling.

[0069] It should be noted that the above is an improvement on central pooling. The pooling process for corner pooling is the same as that for central pooling, and will not be elaborated here.

[0070] Step 4: Based on the center heatmap and corner heatmap, and according to the relationship between the corner points and the center point, model the corner points and the center point to obtain the pointing offset;

[0071] The output of Matching-Head is a 2×H×W shaped pointing offset used to match corner and center points. It includes two channels: the first channel represents the X offset and the second channel represents the Y offset. The H and W of the pointing offset are the same as the width and height of the center heatmap and the corner heatmap.

[0072] By predicting the pointing offset of each corner point on the corner heatmap, the center position of each corner point can be found. The accuracy of the pointing offset prediction directly affects the prediction effect of the model.

[0073] Therefore, to ensure precise matching between corner points and the correct center, a corner-to-center module (CTC) was designed, such as... Figure 3 As shown.

[0074] In step 4, the step of modeling the corner and center points based on the center heatmap and the relationship between the corner and center points to obtain the pointing offset specifically includes:

[0075] Step 401: Input the center heatmap, corner heatmap and basic feature map. Pass the three inputs through a Mixture module to perform max pooling on the input data in the horizontal and vertical directions to obtain the horizontal pooling result and the vertical pooling result, so as to emphasize the row and column where the key point is located.

[0076] Step 402: Concatenate the horizontal pooling results and vertical pooling results with the center heatmap, corner heatmap and basic feature map, and then convolve them to obtain the first output, second output and third output of the Mixture module;

[0077] Step 403: Connect the first output, the second output, and the third output together, and obtain the pointing offset through convolution.

[0078] Heatmaps with rich keypoint information are highly effective in predicting pointing offsets. A lightweight CTC module was also designed, which only accepts basic features as input, eliminating the use of center and corner heatmaps as aids. The basic features are input into the Mixture module, and the output is obtained through convolution. Experimental results show that the lightweight CTC module also performs well in improving matching accuracy.

[0079] Step 5: Based on the pointing offset, a key point matching strategy is adopted to obtain the matching results of corner points and center points by using the pointing offset, and the matching results are output as target detection boxes.

[0080] like Figures 4(a)-4(b) As shown, based on the pointing offset, each corner point can find a corresponding center position. When multiple corner points point to the same center position, there may be multiple objects at this center position.

[0081] To achieve a better match, once a corner point finds its center position, it will continue searching within an n×n range for a suitable center point. If a suitable center point is found, it will also be matched.

[0082] Where n∈[a,b], the specific range depends on the distance from the corner point to the center point according to formula (2).

[0083]

[0084] In the formula, H and W are the output sizes, and x ce The x-coordinate of the center point is represented by x. co Represents the x-coordinate and y-coordinate of the corner point. ce The y-coordinate of the center point is represented by y. co This represents the y-coordinate of the corner point. a and b are hyperparameters used to limit the value of n.

[0085] like Figures 5(a)-5(d) The specific matching strategy and algorithm are as follows.

[0086] (1) Access each channel C in the corner heatmap i .

[0087] (2) Points with a confidence level greater than the threshold at each location are defined as corner points.

[0088] (3) Based on the pointing offset, in each channel C i Find each corner point Corresponding center point

[0089] (4) Add Given the matching range, search for all center points within the range n×n.

[0090] (5) Determine whether the confidence level of the center is greater than the threshold, and retain the angle-center pairs that meet the conditions.

[0091] (6) Among the remaining corner-center pairs, determine whether the corner point is located to the lower right of the center point, and remove the corner-center pairs that do not meet the requirements.

[0092] (7) The filtered corner-center pairs, plus the shared position fine-tuning offset, are output as the target detection box, such as... Figure 6 The output prediction box is shown.

[0093] During the training of the key point target detection model, as a multi-task training process, its loss function is:

[0094] L = L ce +L co +L po +L off (3)

[0095] Among them, L ce Defined as the loss of the central heatmap, L co Defined as the loss of the corner heatmap, L po Defined as the loss for the pointing offset, L off Defined as the loss due to positional offset.

[0096] L ce and L co The formula can be expressed as a variant of focal loss, in the following form:

[0097]

[0098] Where N is the number of center points (corner points), C is the number of categories, H and W are the size of the network output, and p cyx It is the predicted value at position (x,y), g cyx It is the weight at the (x,y) position calculated based on the Gaussian kernel. in, It represents the location of the object's ground-truth keypoints after downsampling. It is the adaptive standard deviation of object size. If two Gaussian functions of the same category overlap, the element-wise maximum value is taken.

[0099] In the formula, α and β are two hyperparameters. In this embodiment, α = 2 and β = 4.

[0100] Apply smoothed L1 loss to calculate L at the truth corner location. po Calculate L at the location of the truth center. off.

[0101]

[0102]

[0103] The model of this invention was compared with the prediction of CenterNet, demonstrating that when key points overlap, the model of this invention can better distinguish and accurately label these objects. Figure 6 Several representative images are shown, in which the center points completely overlap or are very close to each other.

[0104] The experiments of this invention were implemented on the MS-COCO 2017 and PASCALVOC 2007+2012 datasets. On the MS-COCO dataset, training was performed using 115K labeled images from train2017, and validation was performed using 5K images from val2017. On the PASCALVOC dataset, training was performed using 17K labeled images from the entire dataset, and validation was performed using 2K labeled images.

[0105] The impact of camel pooling on existing keypoint-based object detection algorithms was tested. The proposed method was compared with corner pooling in two networks, CornerNet and CentripetalNet, which employ corner pooling. Experimental results under single-scale detection show that this method achieves improved keypoint pooling performance with a small number of parameters. In CenterNet, center pooling and camel pooling were fused. In all experiments, μ = 10, γ = 1.5, and σ = 2 were used in Equation (1). Experimental results also show that the proposed method can achieve better results.

[0106] Table 1. Comparison results of corner pooling in CornerNet and CentripetalNet

[0107]

[0108]

[0109] Table 2 shows the single-scale comparison results of the average accuracy of different modules integrated in CenterNet.

[0110]

[0111] To verify the effectiveness of the CTC module in predicting pointing offsets, we conducted ablation experiments on the matching head in our own network model, demonstrating its effectiveness. Adding the lightweight CTC to the model improved the AP (Average Precision) by 0.9 (from 43.5 to 44.4). Using the CTC module, the model's AP improved by 1.7 (from 43.5 to 45.2). The model's improvement is primarily attributed to its ability to detect large objects. At a single scale, there is still significant room for improvement in the model's performance in detecting small objects.

[0112] Table 3 shows the role of the CTC module in predicting pointing offset, as demonstrated by the ablation experiments.

[0113]

[0114] To address the problem of models failing to distinguish objects when keypoints overlap, this invention proposes a multiple correspondence model. This model uses offsets to point to corner and center pairs, allowing one corner to point to multiple centers, and multiple corners to point to one center. Specific experiments demonstrate that the model of this invention can effectively distinguish targets even when keypoints overlap.

[0115] Example 2

[0116] This embodiment provides a key point target detection system based on multiple correspondence relationships, including:

[0117] The data acquisition module is used to acquire the image data to be detected.

[0118] The object detection module is used to detect objects based on the image data to be detected and the trained keypoint object detection model, and obtain object detection boxes.

[0119] The construction process of the key point target detection model includes: extracting center points and corner points using camel pooling; using the multiple correspondence between center points and corner points, marking the target to be detected using center points and corner points with pointing offsets; wherein the multiple correspondence is: multiple corner points match the same center point or multiple center points match the same corner point.

[0120] Example 3

[0121] This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the key point target detection method based on multiple correspondences as described above.

[0122] Example 4

[0123] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the key point target detection method based on multiple correspondences as described above.

[0124] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0125] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0126] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0127] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0128] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0129] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A key point target detection method based on multiple correspondence relationships, characterized in that, Includes the following steps: Acquire the image data to be detected; Detection is performed based on the image data to be detected and the trained keypoint target detection model to obtain the target detection box; The construction process of the key point target detection model includes: extracting center points and corner points using camel pooling; using the multiple correspondence between center points and corner points, marking the target to be detected using center points and corner points with pointing offsets; wherein the multiple correspondence is: multiple corner points match the same center point or multiple center points match the same corner point. The method of extracting center and corner points using camel hump pooling includes: Pooling is performed in multiple directions, and the pooling results in multiple directions are added together to obtain the emphasis representation of the center point and corner points; Based on the emphasis representation of center point and corner point, combined with the idea of ​​spatial attention, max pooling and average pooling are used along the channel axis to integrate the key point information under different channels into the same space to obtain the integrated result; Based on this integration result, negative center points and corner points around positive key points are suppressed in the horizontal and vertical directions, respectively, resulting in enhanced center points and corner points.

2. The key point target detection method based on multiple correspondences as described in claim 1, characterized in that, After acquiring the image data to be detected, a center heatmap and a corner heatmap are extracted based on the image data to be detected. Based on the center heatmap and the corner heatmap, the corner points and the center points are modeled according to the relationship between the corner points and the center points to obtain the pointing offset of each corner point. Based on the pointing offset of each corner point, a key point matching strategy is adopted. The matching results of the corner point and the center point are obtained by using the offset pointing to the corner point and the center point. The matching results are output as the target detection box.

3. The key point target detection method based on multiple correspondence relationships as described in claim 2, characterized in that, The method, based on the center heatmap and corner heatmap, models the corner and center points according to the relationship between them, obtaining the pointing offset of each corner point, specifically including: Input center heatmap, corner heatmap and basic feature map. Then, pass the three inputs through a Mixture module to perform max pooling on the input data in the horizontal and vertical directions to obtain the horizontal pooling result and the vertical pooling result. The horizontal pooling results and vertical pooling results are concatenated with the center heatmap, corner heatmap and basic feature map, and then convolved to obtain the first output, second output and third output of the Mixture module; The first, second, and third outputs are concatenated together, and the pointing offset is obtained through convolution.

4. The key point target detection method based on multiple correspondence relationships as described in claim 2, characterized in that, The keypoint matching strategy, based on the pointing offset of each corner point, uses the offset pointing to the corner point and the center point to obtain the matching result of the corner point and the center point. Specifically, it includes: Access each channel in the corner heatmap; Based on the pointing offset, find the center point corresponding to each corner point in each channel; Increase the matching range of corner points and reduce the search range. All center points within; Determine if the confidence level of the center is greater than the threshold, and retain the corner-center pairs that meet the conditions; Among the remaining corner-center pairs, those that do not meet the requirements are removed to obtain the filtered corner-center pairs.

5. The key point target detection method based on multiple correspondences as described in claim 1, characterized in that, The key point target detection model also includes a key point position fine-tuning module, in which the prediction of position offset is shared by the center point and the matching corner point of the center point.

6. The key point target detection method based on multiple correspondences as described in claim 1, characterized in that, The loss function of the key point target detection model during training is: in, Defined as the loss of the central heatmap, Defined as the loss of the corner heatmap, Defined as the loss pointing to the offset, Defined as the loss due to positional offset.

7. A key point target detection system based on multiple correspondence relationships, characterized in that, include: The data acquisition module is used to acquire the image data to be detected. The object detection module is used to detect objects based on the image data to be detected and the trained keypoint object detection model, and obtain object detection boxes. The construction process of the key point target detection model includes: extracting center points and corner points using camel pooling; using the multiple correspondence between center points and corner points, marking the target to be detected using center points and corner points with pointing offsets; wherein the multiple correspondence is: multiple corner points match the same center point or multiple center points match the same corner point. The method of extracting center and corner points using camel hump pooling includes: Pooling is performed in multiple directions, and the pooling results in multiple directions are added together to obtain the emphasis representation of the center point and corner points; Based on the emphasis representation of center point and corner point, combined with the idea of ​​spatial attention, max pooling and average pooling are used along the channel axis to integrate the key point information under different channels into the same space to obtain the integrated result; Based on this integration result, negative center points and corner points around positive key points are suppressed in the horizontal and vertical directions, respectively, resulting in enhanced center points and corner points.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the key point target detection method based on multiple correspondences as described in any one of claims 1-6.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the key point target detection method based on multiple correspondences as described in any one of claims 1-6.

Images