Human key point detection method and device, electronic equipment and storage medium

Abstract

The application provides a human key point detection method and device, electronic equipment and storage medium, wherein the human key point detection method comprises: acquiring a to-be-detected human image, the to-be-detected human image being an image collected for a target human in a shielding scene; inputting the to-be-detected human image into a preset key point detection network model to obtain a key point detection result of the target human output by the preset key point detection network model; wherein the preset key point detection network model is used to extract different channel feature maps in the to-be-detected human image, decouple features of the different channel feature maps, respectively perform semantic segmentation on a first shielding object feature map and a first human feature map obtained based on the feature decoupling, perform key point detection on a second shielding object feature map and a second human feature map obtained based on the semantic segmentation, and determine the key point detection result. The application can improve the robustness of human key point detection and improve the application range.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a method, apparatus, electronic device, and storage medium for detecting key points on the human body. Background Technology

[0002] With continuous social progress and technological development, industries such as short videos, live streaming, and online education are emerging, leading to various interactive scenarios. The most common of these is interaction based on human keypoints, and there is a growing demand for interaction based on human keypoints in complex scenarios, particularly in occluded environments. Therefore, improving the detection accuracy and robustness of human keypoints in occluded scenarios is of paramount importance.

[0003] In related technologies, methods for detecting human key points in occluded scenes can either learn human key points by synthesizing occlusion data in the data input space and training the human key point detection network model directly using data augmentation methods, or identify human key points through a multi-stage detection method that first identifies the occlusion location, then detects visible human key points, and finally uses the visible human key points and occlusion location to infer the occluded key points.

[0004] However, since the occlusion data input by data augmentation methods cannot represent rich real occlusion situations, the robustness of human keypoint detection is low. Multi-stage detection methods, which are not end-to-end detection methods, require a large amount of data with occlusion states labeled when identifying occlusion positions. However, currently, data with reasonable and correct occlusion states labeled are scarce and difficult to obtain, thus limiting the applicability of human keypoint detection. Summary of the Invention

[0005] This invention provides a method, apparatus, electronic device, and storage medium for human keypoint detection, which addresses the shortcomings of existing data augmentation methods and multi-stage detection methods for human keypoint detection, namely low robustness and limited applicability. It achieves end-to-end human keypoint detection by decoupling features in the feature extraction space, thereby improving not only the robustness of human keypoint detection but also significantly expanding its applicability.

[0006] This invention provides a method for detecting key points on the human body, comprising:

[0007] Acquire an image of the human body to be detected, wherein the image of the human body to be detected is an image of a target human body in an occluded scene;

[0008] The human image to be detected is input into a preset key point detection network model to obtain the key point detection results of the target human body output by the preset key point detection network model;

[0009] The preset key point detection network model is used to extract different channel feature maps from the human image to be detected, and to decouple the different channel feature maps. Based on the first occlusion feature map and the first human feature map obtained by the feature decoupling, semantic segmentation is performed respectively. Based on the second occlusion feature map and the second human feature map obtained by the semantic segmentation, key point detection is performed to determine the key point detection result.

[0010] According to a method for detecting human key points provided by the present invention, the step of extracting different channel feature maps from the human image to be detected includes:

[0011] Based on a preset spatial regularization loss function, different local spatial locations in the human body image to be detected are mapped to different channels to determine different channel feature maps in the human body image to be detected; the preset spatial regularization loss function is used to limit the uncorrelation between convolutional kernels and the uncorrelation between channels.

[0012] According to a method for detecting key points in the human body provided by the present invention, the feature decoupling of the different channel feature maps includes:

[0013] The different channel feature maps are weighted based on different preset occlusion channel weights;

[0014] The different channel feature maps are weighted based on different preset human body channel weights.

[0015] According to the present invention, a method for detecting human key points includes performing semantic segmentation on the first occlusion feature map and the first human feature map obtained based on the feature decoupling, respectively:

[0016] The first occluder feature map and the first human body feature map obtained by decoupling the features are downsampled to determine the target occluder feature map and the target human body feature map.

[0017] Semantic segmentation is performed on the feature map of the target occluder and the feature map of the target human body, respectively.

[0018] According to the present invention, a method for detecting key points on the human body includes a training process for a preset key point detection network model comprising:

[0019] A sample human body image and an initial keypoint detection network model are determined. The initial keypoint detection network model includes a feature extraction model, a feature decoupling model, an occlusion semantic segmentation model, a human body semantic segmentation model, and a keypoint detection model.

[0020] The sample human body image is input into the initial keypoint detection network model for training of feature extraction, feature decoupling, downsampling, semantic segmentation, and keypoint detection, thereby determining the preset keypoint detection network model; wherein, the feature decoupling includes weighting the different sample channel feature maps obtained by feature extraction using different sample occlusion channel weights and different sample human body channel weights respectively.

[0021] According to a method for detecting human key points provided by the present invention, the step of inputting the sample human image into an initial key point detection network model for training of feature extraction, feature decoupling, downsampling, semantic segmentation, and key point detection, and determining the preset key point detection network model, includes:

[0022] The sample human body image is input into the initial key point detection network model for iterative training of feature extraction, feature decoupling, downsampling, semantic segmentation and key point detection. The sample occlusion segmentation result and sample human body segmentation result obtained after training for a preset number of iterations, as well as the key point detection result of the intermediate key point detection network model obtained after training for a preset number of iterations, are obtained.

[0023] Determine whether the occlusion segmentation index value of the sample occlusion segmentation result reaches the preset occlusion segmentation index threshold, whether the human body segmentation index value of the sample human body segmentation result reaches the preset human body segmentation index threshold, and whether the detection index value of the key point detection result reaches the preset detection threshold.

[0024] When the occlusion segmentation index value reaches the preset occlusion segmentation index threshold, the human body segmentation index value reaches the preset human body segmentation index threshold, and the detection index value reaches the preset detection threshold, training is stopped, and the intermediate key point detection network model corresponding to the training stop is determined as the preset key point detection network model.

[0025] When it is determined that the detection index value does not reach the preset detection threshold, the occlusion segmentation index value does not reach the preset occlusion segmentation index threshold, and / or the human body segmentation index value does not reach the preset human body segmentation index threshold, the parameters of the intermediate key point detection network model are updated, and the sample occlusion channel weights and / or sample human body channel weights participating in the training for the preset number of iterations are adjusted. Based on the adjusted sample occlusion channel weights, the adjusted sample human body channel weights, and / or the updated model parameters, the intermediate key point detection network model is iterated and trained again; until training stops, the preset key point detection network model is determined.

[0026] According to the present invention, a method for detecting human key points includes a feature decoupling model comprising an occlusion channel attention model and a human body channel attention model. The method involves iteratively training the sample human image into the initial key point detection network model for feature extraction, feature decoupling, downsampling, semantic segmentation, and key point detection, comprising:

[0027] For each training iteration, the sample human body image is input into the feature extraction model of the initial keypoint detection network model to obtain different sample channel feature maps in the sample human body image;

[0028] The different sample channel feature maps are input into the occlusion channel attention model for convolution, downsampling and full connection to determine the different sample occlusion channel weights;

[0029] The different sample channel feature maps are input into the human body channel attention model and then subjected to convolution, downsampling and full connection to determine the different sample human body channel weights.

[0030] The different sample occlusion channel weights and the different sample human body channel weights are used to weight the different sample channel feature maps respectively to determine the first sample occlusion feature map and the first sample human body feature map.

[0031] The first occluder feature map and the first human body feature map of the sample are downsampled respectively to determine the target occluder feature map and the target human body feature map of the sample.

[0032] Semantic segmentation is performed on the sample target occlusion feature map and the sample target human body feature map to determine the sample second occlusion feature map and the sample second human body feature map;

[0033] Key point detection is performed on the second occlusion feature map and the second human feature map of the sample.

[0034] According to a method for detecting key points on the human body provided by the present invention, the method further includes:

[0035] While determining that the intermediate key point detection network model corresponding to the training stop is the preset key point detection network model, the adjusted sample occlusion channel weights corresponding to the training stop are determined to be different preset occlusion channel weights, and the adjusted sample human body channel weights corresponding to the training stop are determined to be different preset human body channel weights.

[0036] The present invention also provides a human body key point detection device, comprising:

[0037] The acquisition module is used to acquire the human body image to be detected, which is an image of the target human body in an occluded scene;

[0038] The detection module is used to input the human image to be detected into a preset key point detection network model to obtain the key point detection results of the target human body output by the preset key point detection network model;

[0039] The preset key point detection network model is used to extract different channel feature maps from the human image to be detected, and to decouple the different channel feature maps. Based on the first occlusion feature map and the first human feature map obtained by the feature decoupling, semantic segmentation is performed respectively. Based on the second occlusion feature map and the second human feature map obtained by the semantic segmentation, key point detection is performed to determine the key point detection result.

[0040] The present invention also provides an electronic device, including 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 human body key point detection method as described above.

[0041] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the human key point detection method as described above.

[0042] The present invention provides a method, apparatus, electronic device, and storage medium for human keypoint detection. The method involves inputting a human image to be detected into a preset keypoint detection network model. First, different channel feature maps are extracted from the human image. Then, feature decoupling is performed on the different channel feature maps. Next, semantic segmentation is performed on the first occlusion feature map and the first human feature map obtained from feature decoupling. Finally, keypoint detection is performed on the second occlusion feature map and the second human feature map obtained from semantic segmentation. This method accurately and quickly achieves end-to-end human keypoint detection with feature decoupling in the feature extraction space. Compared to multi-stage methods that first identify occlusion and then process it, this method has the advantages of simple training and does not require information on keypoint occlusion state annotation. It not only improves the robustness of human keypoint detection but also significantly expands its applicability, thereby greatly improving the accuracy of human keypoint detection in occluded scenarios. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0044] Figure 1 This is a flowchart illustrating the human body key point detection method provided by the present invention;

[0045] Figure 2 This is a schematic diagram of the human body key point detection device provided by the present invention;

[0046] Figure 3 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0048] The following is combined Figures 1-3 This invention describes a method, apparatus, electronic device, and storage medium for detecting human key points. The execution entity of the human key point detection method can be a terminal device or a server. The terminal device can be a personal computer (PC), portable device, laptop, smartphone, tablet computer, portable wearable device, or other electronic device. The server can be a standalone server or a server cluster composed of multiple servers. For example, the server can be a physical server containing independent hosts, a virtual server hosted by a host cluster, a cloud server, etc. This invention does not limit the specific form of the terminal device, nor does it specifically limit the specific form of the server.

[0049] It should be noted that the following method embodiments are described using a terminal device as the execution subject, and the execution subject of the following method embodiments can be part or all of the terminal device.

[0050] Figure 1 This is a flowchart illustrating the human key point detection method provided in an embodiment of the present invention, as shown below. Figure 1 As shown, this method for detecting key points on the human body includes the following steps:

[0051] Step 110: Obtain the human body image to be detected. The human body image to be detected is an image of the target human body in an occluded scene.

[0052] The target human body can be at least one, and the occlusion scene is a scene where at least one key point of the target human body is occluded by other objects, such as tables or chairs. For example, an occlusion scene is a scene where part of the target human body's right leg is occluded by the chair when the target human body is standing behind it.

[0053] Specifically, the terminal device acquires the image of the human body to be detected in several ways: by using its built-in camera to capture the target human body in an occluded scene; by instructing a separate camera connected to the terminal device to capture the target human body in an occluded scene; or by selecting images of the target human body captured in an occluded scene from a pre-stored image dataset. No specific limitations are placed on the method by which the terminal device acquires the image of the human body to be detected.

[0054] Step 120: Input the human image to be detected into the preset key point detection network model to obtain the key point detection results of the target human body output by the preset key point detection network model;

[0055] The preset key point detection network model is used to extract different channel feature maps in the human body image to be detected, and to decouple the different channel feature maps. Based on the first occlusion feature map and the first human body feature map obtained by feature decoupling, semantic segmentation is performed respectively. Based on the second occlusion feature map and the second human body feature map obtained by semantic segmentation, key point detection is performed to determine the key point detection result.

[0056] Specifically, when the target human body is in an occluded scene, at least one key point of the target human body will be occluded. Occlusion leads to feature nesting and entanglement. Therefore, in order to reduce the dependence on the diversity of occlusion data, this invention does not add occlusion data at the data input end. Instead, it decouples the extracted multiple channel feature maps in the feature space. That is, in the feature extraction space, multiple channel feature maps are explicitly decoupled into a first occluder feature map and a first human body feature map. By performing semantic segmentation on the first occluder feature map and the first human body feature map respectively, their sizes are converted into a second occluder feature map and a second human body feature map that are suitable for the accuracy and precision requirements of subsequent human body key point detection. At this time, key point detection is performed on the second occluder feature map and the second human body feature map to obtain rich and accurate human body key point detection results.

[0057] The human keypoint detection method provided by this invention involves a terminal device inputting a human image to be detected into a preset keypoint detection network model. First, different channel feature maps are extracted from the human image. Then, feature decoupling is performed on the different channel feature maps. Next, semantic segmentation is performed on the first occlusion feature map and the first human feature map obtained from feature decoupling. Finally, keypoint detection is performed on the second occlusion feature map and the second human feature map obtained from semantic segmentation. This method accurately and quickly achieves end-to-end human keypoint detection with feature decoupling in the feature extraction space. Compared to multi-stage methods that first identify occlusion and then process it, this method has the advantages of simple training and does not require information on keypoint occlusion status annotation. It not only improves the robustness of human keypoint detection but also significantly expands its applicability, thereby greatly improving the accuracy of human keypoint detection in occluded scenarios.

[0058] Optionally, different channel feature maps can be extracted from the human image to be detected. The implementation process may include:

[0059] Based on the preset spatial regularization loss function, different local spatial locations in the human body image to be detected are mapped to different channels to determine the feature maps of different channels in the human body image to be detected; the preset spatial regularization loss function is used to limit the uncorrelation between convolution kernels and the uncorrelation between channels.

[0060] Specifically, when the preset keypoint detection network model is a network model trained by adding feature decoupling and a preset spatial regularization loss function between the feature extraction and downsampling stages in the existing residual network, and adding semantic segmentation after the downsampling stage, regularization constraints, i.e., preset spatial regularization loss functions, can be pre-set in the feature space. When the terminal device uses the preset keypoint detection network model for feature extraction, it can use the preset spatial regularization loss function, pre-established and stored in the feature space, to map different local spatial locations in the human image to be detected to different channels, thus determining different channel feature maps in the human image. This ensures that each local spatial location in the human image corresponds to a channel feature map, and each channel feature map responds to different local spatial locations. In other words, a portion of the channel feature maps represents the features of the occluded object, while the remaining portion represents the features of the occluded target human body. Furthermore, the preset spatial regularization loss function includes a first loss function that penalizes the correlation of convolution kernels and a second loss function that penalizes the correlation of channel responses. The first loss function is used to constrain each convolution kernel to be as uncorrelated as possible and to activate different image modes and local spatial locations. The second loss function is used to constrain each channel to be as uncorrelated as possible. The calculation formulas for the first and second loss functions are as follows:

[0061]

[0062] in, This represents the value at the p-th position in the parameters of the i-th convolution kernel. Let r be the value at the p-th position of the j-th convolution kernel parameters. i For the feature map of the i-th channel corresponding to the parameters of the i-th convolutional kernel, r j Let F be the feature map of the j-th channel corresponding to the parameters of the j-th convolution kernel. The subscript F represents the F-norm operation, L1 is the first loss function, and L2 is the second loss function.

[0063] The human key point detection method provided by this invention determines different channel feature maps in the human image to be detected by adding a preset spatial regularization loss function to the feature extraction space to limit the correlation between convolution kernels and between channels. This achieves the purpose of each channel feature map representing different local spatial locations in the human image to be detected, laying the foundation for accurate and efficient feature decoupling in the future.

[0064] Optionally, feature decoupling can be performed on different channel feature maps, and the implementation process may include:

[0065] Based on different preset occlusion channel weights, different channel feature maps are weighted; based on different preset human body channel weights, different channel feature maps are weighted.

[0066] Specifically, when using a preset keypoint detection network model for feature decoupling, the feature decoupling purpose can be achieved by using different preset occlusion channel weights and different preset human body channel weights that have been learned and stored in advance to weight different channel feature maps respectively. The number of preset occlusion channel weights is the same as the number of channel feature maps and corresponds one-to-one, and the number of preset human body channel weights is the same as the number of channel feature maps and corresponds one-to-one.

[0067] The human key point detection method provided by the present invention achieves the purpose of occlusion decoupling in the feature space by using multiple preset human channel weights and weighting multiple channel feature maps respectively by the terminal device, thereby improving the reliability and applicability of feature decoupling.

[0068] Optionally, semantic segmentation can be performed on the first occlusion feature map and the first human body feature map obtained from feature decoupling, and the implementation process may include:

[0069] First, the first occluder feature map and the first human body feature map obtained by feature decoupling are downsampled to determine the target occluder feature map and the target human body feature map; then, semantic segmentation is performed on the target occluder feature map and the target human body feature map respectively.

[0070] Among them, the number of first occlusion feature maps is at least one, the number of first human body feature maps is at least one, and the sum of the number of first occlusion feature maps and first human body feature maps is equal to the total number of channel feature maps.

[0071] Specifically, the preset keypoint detection network model can be a network model trained by adding feature decoupling and a preset spatial regularization loss function between the feature extraction and downsampling stages in the existing residual network, and adding semantic segmentation after the downsampling stage. The residual network can be ResNet50, VGG, MobileNet, etc. The preset spatial regularization loss function can also be set in the feature extraction stage of the existing residual network. For example, when the residual network is ResNet50, the preset spatial regularization loss function can be set in the third stage of feature extraction in ResNet50, or between the third stage of feature extraction and the fourth stage of downsampling in ResNet50. No specific limitations are made here. Based on this, after downsampling the first occlusion feature map and the first human feature map obtained from feature decoupling, semantic segmentation is then performed on the target occlusion feature map and the target human feature map obtained after downsampling, to ensure that the target occlusion feature map and the target human feature map are subsequently processed into a second occlusion feature map and a second human feature map that can improve the efficiency of human keypoint detection.

[0072] The human body key point detection method provided by the present invention enables the terminal device to quickly and accurately generate a second occlusion feature map and a second human body feature map by first downsampling and then performing semantic segmentation on the first occlusion feature map and the first human body feature map obtained by feature decoupling, thereby ensuring the comprehensiveness and reliability of subsequent human body key point detection.

[0073] Optionally, the training process of the preset keypoint detection network model includes:

[0074] First, the sample human images and the initial keypoint detection network model are determined. The initial keypoint detection network model includes a feature extraction model, a feature decoupling model, an occlusion semantic segmentation model, a human semantic segmentation model, and a keypoint detection model. Then, the sample human images are input into the initial keypoint detection network model for training of feature extraction, feature decoupling, downsampling, semantic segmentation, and keypoint detection to determine the preset keypoint detection network model.

[0075] Feature decoupling involves weighting the different sample channel feature maps obtained from feature extraction using different sample occlusion channel weights and different sample human body channel weights. The sample human body images can be images collected from human bodies in occluded scenes; the sample human body can be the same as or different from the target human body. There is no limitation here. The number of sample human body images can be multiple. The specific structure of the keypoint detection model is not specifically limited; for example, the keypoint detection model can be set to two or more convolutional kernels.

[0076] Specifically, the initial keypoint detection network model can be a network model obtained by adding feature decoupling and a preset spatial regularization loss function between the feature extraction and downsampling stages in the existing residual network, and adding a semantic segmentation stage after the downsampling stage. Therefore, the initial keypoint detection network model can be determined to include a feature extraction model, a feature decoupling model, an occlusion semantic segmentation model, a human semantic segmentation model, and a keypoint detection model. The feature decoupling model can include decoupling the occlusion channel feature map representing the occlusion and the human channel feature map representing the human body sample from the different sample channel feature maps obtained from feature extraction. The different sample occlusion channel weights and different sample human body channel weights used in the feature decoupling process also need to be adjusted during model training. The initial values ​​of the sample occlusion channel weights and sample human body channel weights can be determined by random number assignment or by learning different sample channel feature maps using a self-attention mechanism. No specific limitations are made here.

[0077] Based on this, the initial keypoint detection network model is trained using sample human images for feature extraction, feature decoupling, downsampling, semantic segmentation, and keypoint detection. For example, the number of training iterations can be set. After training the initial keypoint detection network model with a set number of sample human images, the output keypoint detection results after the set number of training iterations can be sent to the client so that the client can determine whether the accuracy and comprehensiveness of the output keypoint detection results meet the actual needs.

[0078] When the terminal device receives the first feedback from the client, indicating that the accuracy and comprehensiveness of the currently output keypoint detection results meet the actual needs, it can use the corresponding keypoint detection network model as the preset keypoint detection network model. Conversely, when the terminal device receives the second feedback from the client, indicating that the accuracy and comprehensiveness of the currently output keypoint detection results do not meet the actual needs and the training iteration count needs adjustment, it can adjust the image order of multiple sample human images, use the keypoint detection network model corresponding to the output keypoint detection results as the new initial keypoint detection network model, adjust the training iteration count, and then train the new initial keypoint detection network model for the adjusted number of training iterations based on the adjusted sample human images and the adjusted training iteration count. This process continues until the first feedback from the client is received.

[0079] The human keypoint detection method provided by this invention determines the preset keypoint detection network model by training the initial keypoint detection network model using sample human images through feature extraction, feature decoupling, downsampling, semantic segmentation, and keypoint detection. This achieves the purpose of training the network model based on sample human images, avoiding the problem of low robustness of keypoint detection caused by using human keypoint detection network models for training when occluded data is input at the input end in traditional methods, and greatly improving the robustness of keypoint detection.

[0080] Optionally, the sample human body image is input into an initial keypoint detection network model for training, including feature extraction, feature decoupling, downsampling, semantic segmentation, and keypoint detection. This determines the preset keypoint detection network model, the implementation of which may include:

[0081] First, the sample human images are input into an initial keypoint detection network model for iterative training of feature extraction, feature decoupling, downsampling, semantic segmentation, and keypoint detection. This yields the occlusion segmentation results and human body segmentation results obtained after a preset number of iterations, as well as the keypoint detection results from an intermediate keypoint detection network model after the same preset number of iterations. Next, it is determined whether the occlusion segmentation index value of the sample occlusion segmentation results reaches a preset occlusion segmentation index threshold, the human body segmentation index value of the sample human body segmentation results reaches a preset human body segmentation index threshold, and the keypoint detection index value of the keypoint detection results reaches a preset detection threshold. Finally, it is confirmed that the occlusion segmentation index value and the human body segmentation index value have reached the preset human body segmentation index thresholds. When the detection index value and the detection index value reach the preset detection threshold, training stops, and the intermediate keypoint detection network model corresponding to the training stop is determined as the preset keypoint detection network model. Conversely, when it is determined that the detection index value, the occlusion segmentation index value, and / or the human segmentation index value do not reach the preset human segmentation index threshold, the parameters of the intermediate keypoint detection network model are updated, and the occlusion channel weights and / or human channel weights of the samples participating in the training for a preset number of iterations are adjusted. Based on the adjusted sample occlusion channel weights, the adjusted sample human channel weights, and / or the updated model parameters, the intermediate keypoint detection network model is iterated and trained again until the training stops and the preset keypoint detection network model is determined.

[0082] The preset number of iterations can be one or more, without specific limitations. The occlusion segmentation metric can be the occlusion loss value or mean squared error value of the sample occlusion segmentation results obtained after training for the preset number of iterations. The human segmentation metric can be the loss value or mean squared error value of the sample human segmentation results obtained after training for the preset number of iterations. Correspondingly, the preset occlusion segmentation metric threshold can be the occlusion segmentation loss threshold or the occlusion segmentation mean squared error threshold, and the preset human segmentation metric threshold can also be the human segmentation loss threshold or the human segmentation mean squared error threshold. Completing the preset number of iterations of training is considered as completing one round of training.

[0083] Specifically, when the occlusion segmentation index value reaches the preset occlusion segmentation index threshold, the human body segmentation index value reaches the preset human body segmentation index threshold, and the detection index value reaches the preset detection threshold, the terminal device can determine that the keypoint detection network model trained after a preset number of iterations is a well-trained keypoint detection network model. Conversely, when the occlusion segmentation index value does not reach the preset occlusion segmentation index threshold, the human body segmentation index value does not reach the preset human body segmentation index threshold, and / or the detection index value does not reach the preset detection threshold, the terminal device can determine that the keypoint detection network model trained after a preset number of iterations is not a well-trained network model. At this time, the parameters of the intermediate keypoint detection network model obtained after this round of training can be updated, and the weights of the sample occlusion channel and / or the sample human body channel participating in the training of the preset number of iterations can be adjusted. Based on the adjusted sample occlusion channel weights, the adjusted sample human body channel weights, and / or the updated model parameters, the intermediate keypoint detection network model is iteratively trained again. The input of the intermediate keypoint detection network model with updated model parameters can be a new sample human body image or a sample human body image from the previous training round.

[0084] It should be noted that during the model training process of this invention, a multi-task learning framework related to human keypoint detection can be introduced for supervised learning. This multi-task learning framework can include two segmentation tasks, namely, introducing occlusion semantic segmentation and human semantic segmentation tasks for supervised learning feature decoupling. The human semantic segmentation task can also be replaced by tasks such as human detection and human instance segmentation. No specific limitations are made here. This achieves the goal of incorporating more human-related datasets using a multi-task framework, reducing the dependence on the diversity of occlusion patterns in the human keypoint training data. Furthermore, during model training, the outputs of the occlusion semantic segmentation model and the human semantic segmentation model are used as supervisory signals for supervised learning. That is, the sample occlusion segmentation results and sample human segmentation results are used to supervise the learning of different sample occlusion channel weights and different sample human channel weights until the metrics of both segmentation results meet the standards. The occlusion semantic segmentation model and the human semantic segmentation model can adopt existing segmentation networks. This invention does not specifically limit the specific structure of the segmentation network and the weights cannot be shared.

[0085] The human keypoint detection method provided by this invention determines whether the intermediate keypoint detection network model is a pre-trained preset keypoint detection network model, or whether it is retrained after adjusting the corresponding channel weights and / or updating the model parameters, based on the sample human image iteratively trained on an initial keypoint detection network model. This process is repeated until the preset keypoint detection network model is obtained. This improves the flexibility of training the network model, thereby enhancing the reliability of the preset keypoint detection network model and ensuring improved robustness and accuracy of human keypoint detection in subsequent practical applications.

[0086] Optionally, the feature decoupling model includes an occlusion channel attention model and a human body channel attention model. Sample human images are input into an initial keypoint detection network model for iterative training of feature extraction, feature decoupling, downsampling, semantic segmentation, and keypoint detection. The implementation process may include:

[0087] For each training iteration, the sample human images are input into the feature extraction model of the initial keypoint detection network model to obtain different sample channel feature maps in the sample human images. These different sample channel feature maps are then input into the occlusion channel attention model for convolution, downsampling, and full connection to determine different sample occlusion channel weights. Simultaneously, these different sample channel feature maps are input into the human channel attention model for convolution, downsampling, and full connection to determine different sample human channel weights. Then, the different sample channel feature maps are weighted using the different sample occlusion channel weights and the different sample human channel weights to determine the first sample occlusion feature map and the first sample human feature map. Further downsampling is performed on the first sample occlusion feature map and the first sample human feature map to determine the sample target occlusion feature map and the sample target human feature map. Semantic segmentation is then performed on the sample target occlusion feature map and the sample target human feature map to determine the second sample occlusion feature map and the second sample human feature map. Finally, keypoint detection is performed on the second sample occlusion feature map and the second sample human feature map.

[0088] Specifically, different attention weights are set in the occlusion channel attention model and the human channel attention model. That is, different sample occlusion channel weights are set in the occlusion channel attention model, and different sample human channel weights are set in the human channel attention model. These different sample occlusion channel weights are determined by performing convolution, downsampling, and fully connected operations on different sample channel feature maps using the occlusion channel attention model. The number of sample occlusion channel weights corresponds one-to-one with the number of sample channel feature maps. Similarly, different sample human channel weights are determined by performing convolution, downsampling, and fully connected operations on different sample channel feature maps using the human channel attention model. The number of sample human channel weights also corresponds one-to-one with the number of sample channel feature maps. Based on this, different sample occlusion channel weights can be correlated with different sample channel feature maps. Figure 1 After multiplication, a new sample occlusion feature map is obtained, which is multiple sample first occlusion feature maps. The number of sample first occlusion feature maps is the same as the number of sample occlusion channel weights. At the same time, different sample human body channel weights are multiplied by different sample channel feature maps. Figure 1 After multiplication, a new sample human body feature map is obtained, which is multiple sample first human body feature maps. The number of sample first human body feature maps is the same as the number of sample human body channel weights, thereby achieving the purpose of feature decoupling. The sample first occlusion feature map and sample first human body feature map output by the occlusion channel attention model and the human body channel attention model are downsampled and semantically segmented in sequence to determine multiple sample second occlusion feature maps and multiple sample second human body feature maps. At this time, the multiple sample second occlusion feature maps and multiple sample second human body feature maps can be superimposed to perform human body key point detection.

[0089] It should be noted that when the network structures of both the occlusion semantic segmentation model and the human semantic segmentation model include 3 upsampling layers, 3 convolutional layers, 3 batch normalization layers, and 3 ReLU activation layers, each sample target occlusion feature map obtained after downsampling undergoes 3 rounds of processing, with each round including upsampling, convolution, batch normalization, and ReLU activation. Similarly, each sample target human feature map obtained after downsampling also undergoes 3 rounds of processing, with each round including upsampling, convolution, batch normalization, and ReLU activation. In this way, each sample target occlusion feature map is processed into a sample second occlusion feature map that can improve the robustness of keypoint detection, and each sample target human feature map is processed into a sample second human feature map that can improve the robustness of keypoint detection. For example, when the upsampling rate is 3 and the convolution size is 3*3 each time, if both the sample target human body feature map and the sample target human body feature map are 8*8, then the size of the sample second occlusion feature map obtained after processing the sample target human body feature map and the sample target human body feature map for 3 rounds is 64*64 and the size of the sample second human body feature map is also 64*64.

[0090] The human keypoint detection method provided by this invention involves a terminal device first determining the weights of different sample occlusion channels and different sample human channels using an occlusion channel attention model and a human channel attention model. Then, the corresponding weights are weighted to determine the first occlusion feature map and the first human feature map of the sample. Next, downsampling, semantic segmentation, and keypoint detection are performed on the first occlusion feature map and the first human feature map of the sample in sequence. This achieves the purpose of one-time iterative training of feature extraction, feature decoupling, downsampling, semantic segmentation, and keypoint detection of the sample human image, thereby improving the accuracy and reliability of model training.

[0091] Optionally, the method of the present invention may further include:

[0092] While determining that the intermediate keypoint detection network model corresponding to the training stop is the preset keypoint detection network model, the adjusted sample occlusion channel weights corresponding to the training stop are determined to be different preset occlusion channel weights, and the adjusted sample human body channel weights corresponding to the training stop are determined to be different preset human body channel weights.

[0093] Specifically, when the initial keypoint detection network model is trained to convergence and the loss accuracy meets the preset accuracy requirements, the occlusion channel weights and human body channel weights required in the feature decoupling process are also adjusted to meet the preset requirements through supervised learning. Furthermore, the occlusion channel weights and human body channel weights required in the feature decoupling process can be learned to meet the preset requirements simultaneously or sequentially, without specific limitations here.

[0094] The human key point detection method provided by this invention allows the terminal device to determine the channel weights of different preset occlusion objects and different preset human body channels to meet preset requirements while determining the trained preset key point detection network model. This enables the terminal device to quickly and accurately obtain human key point detection in practical applications, thereby improving the detection accuracy and robustness of human key point detection in images of target human bodies in occluded scenes.

[0095] The human body key point detection device provided by the present invention is described below. The human body key point detection device described below can be referred to in correspondence with the human body key point detection method described above.

[0096] Reference Figure 2 The diagram below shows the structure of the human body key point detection device provided by the present invention. Figure 2 As shown, the human body key point detection device 200 includes:

[0097] The acquisition module 210 is used to acquire the human body image to be detected, which is an image of the target human body in an occluded scene.

[0098] The detection module 220 is used to input the human image to be detected into a preset keypoint detection network model to obtain the keypoint detection result of the target human body output by the preset keypoint detection network model. The preset keypoint detection network model is used to extract different channel feature maps in the human image to be detected, and to decouple the different channel feature maps. Based on the first occlusion feature map and the first human body feature map obtained by feature decoupling, semantic segmentation is performed respectively. Based on the second occlusion feature map and the second human body feature map obtained by semantic segmentation, keypoint detection is performed to determine the keypoint detection result.

[0099] Optionally, the detection module 220 can be used to map different local spatial locations in the human image to be detected to different channels based on a preset spatial regularization loss function, thereby determining different channel feature maps in the human image to be detected; the preset spatial regularization loss function is used to limit the uncorrelation between convolutional kernels and the uncorrelation between channels.

[0100] Optionally, the detection module 220 can also be used to weight different channel feature maps based on different preset occlusion channel weights; and to weight different channel feature maps based on different preset human body channel weights.

[0101] Optionally, the detection module 220 can also be used to downsample the first occlusion feature map and the first human body feature map obtained by feature decoupling to determine the target occlusion feature map and the target human body feature map; and to perform semantic segmentation on the target occlusion feature map and the target human body feature map respectively.

[0102] Optionally, the device of the present invention may further include a training module for training and generating a preset keypoint detection network model. The training process of the preset keypoint detection network model includes: determining a sample human image and an initial keypoint detection network model, wherein the initial keypoint detection network model includes a feature extraction model, a feature decoupling model, an occlusion semantic segmentation model, a human semantic segmentation model, and a keypoint detection model; inputting the sample human image into the initial keypoint detection network model for training of feature extraction, feature decoupling, downsampling, semantic segmentation, and keypoint detection, thereby determining the preset keypoint detection network model; wherein, feature decoupling includes weighting the different sample channel feature maps obtained by feature extraction using different sample occlusion channel weights and different sample human channel weights respectively.

[0103] Optionally, the training module can also be used to input sample human images into an initial keypoint detection network model for iterative training of feature extraction, feature decoupling, downsampling, semantic segmentation, and keypoint detection. This allows for obtaining the sample occlusion segmentation results and sample human segmentation results after a preset number of iterations, as well as the keypoint detection results of the intermediate keypoint detection network model after a preset number of iterations. It can also determine whether the occlusion segmentation index value of the sample occlusion segmentation results reaches a preset occlusion segmentation index threshold, whether the human segmentation index value of the sample human segmentation results reaches a preset human segmentation index threshold, and whether the keypoint detection index value reaches a preset detection threshold. Finally, it can determine whether the occlusion segmentation index value reaches the preset occlusion segmentation index threshold and the human segmentation index value reaches the preset human segmentation index threshold. When the body segmentation threshold and the detection index value reach the preset detection threshold, training stops, and the intermediate keypoint detection network model corresponding to the training stop is determined as the preset keypoint detection network model. When the detection index value, the occlusion segmentation index value, and / or the human segmentation index value do not reach the preset human segmentation index threshold, the parameters of the intermediate keypoint detection network model are updated, and the occlusion channel weights and / or human channel weights of the samples participating in the training for a preset number of iterations are adjusted. The intermediate keypoint detection network model updated based on the adjusted occlusion channel weights, the adjusted human channel weights, and / or model parameters is iterated and trained again until the training stops and the preset keypoint detection network model is determined.

[0104] Optionally, the training module can also be used for each training iteration to input sample human images into the feature extraction model of the initial keypoint detection network model to obtain different sample channel feature maps in the sample human images; input the different sample channel feature maps into the occlusion channel attention model for convolution, downsampling, and full connection to determine different sample occlusion channel weights; input the different sample channel feature maps into the human channel attention model for convolution, downsampling, and full connection to determine different sample human channel weights; use the different sample occlusion channel weights and the different sample human channel weights to weight the different sample channel feature maps respectively to determine the first sample occlusion feature map and the first sample human feature map; downsample the first sample occlusion feature map and the first sample human feature map respectively to determine the sample target occlusion feature map and the sample target human feature map; perform semantic segmentation on the sample target occlusion feature map and the sample target human feature map respectively to determine the second sample occlusion feature map and the second sample human feature map; and perform keypoint detection on the second sample occlusion feature map and the second sample human feature map.

[0105] Optionally, the training module can also be used to determine, at the same time, the intermediate keypoint detection network model corresponding to the training stop as the preset keypoint detection network model, the adjusted sample occlusion channel weights corresponding to the training stop as different preset occlusion channel weights, and the adjusted sample human body channel weights corresponding to the training stop as different preset human body channel weights.

[0106] The human keypoint detection device provided by this invention first extracts different channel feature maps from the human image to be detected by inputting it into a preset keypoint detection network model. Then, it performs feature decoupling on the different channel feature maps. Next, it performs semantic segmentation on the first occlusion feature map and the first human feature map obtained by feature decoupling. Finally, it performs keypoint detection on the second occlusion feature map and the second human feature map obtained by semantic segmentation. This method accurately and quickly achieves the end-to-end human keypoint detection objective by performing feature decoupling in the feature extraction space. Compared with the multi-stage method of first identifying occlusion and then processing occlusion, it has the advantages of simple training and no need for keypoint occlusion state annotation information. It not only improves the robustness of human keypoint detection, but also greatly expands the scope of application, thereby greatly improving the accuracy of human keypoint detection in occluded scenarios.

[0107] Figure 3 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 3 As shown, the electronic device 300 may include: a processor 310, a communication interface 320, a memory 330, and a communication bus 340, wherein the processor 310, the communication interface 320, and the memory 330 communicate with each other via the communication bus 340. The processor 310 can call logical instructions in the memory 330 to execute a human key point detection method, which includes:

[0108] Acquire the human body image to be detected, which is an image of the target human body in an occluded scene;

[0109] The human image to be detected is input into the preset key point detection network model to obtain the key point detection results of the target human body output by the preset key point detection network model.

[0110] The preset key point detection network model is used to extract different channel feature maps in the human body image to be detected, and to decouple the different channel feature maps. Based on the first occlusion feature map and the first human body feature map obtained by feature decoupling, semantic segmentation is performed respectively. Based on the second occlusion feature map and the second human body feature map obtained by semantic segmentation, key point detection is performed to determine the key point detection result.

[0111] Furthermore, the logical instructions in the aforementioned memory 330 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present 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 described in the various embodiments of the present 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.

[0112] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program that can be stored on a non-transitory computer-readable storage medium, wherein when the computer program is executed by a processor, the computer is able to execute the human key point detection method provided by the above methods, the method comprising:

[0113] Acquire the human body image to be detected, which is an image of the target human body in an occluded scene;

[0114] The human image to be detected is input into the preset key point detection network model to obtain the key point detection results of the target human body output by the preset key point detection network model.

[0115] The preset key point detection network model is used to extract different channel feature maps in the human body image to be detected, and to decouple the different channel feature maps. Based on the first occlusion feature map and the first human body feature map obtained by feature decoupling, semantic segmentation is performed respectively. Based on the second occlusion feature map and the second human body feature map obtained by semantic segmentation, key point detection is performed to determine the key point detection result.

[0116] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the human key point detection method provided by the methods described above, the method comprising:

[0117] Acquire the human body image to be detected, which is an image of the target human body in an occluded scene;

[0118] The human image to be detected is input into the preset key point detection network model to obtain the key point detection results of the target human body output by the preset key point detection network model.

[0119] The preset key point detection network model is used to extract different channel feature maps in the human body image to be detected, and to decouple the different channel feature maps. Based on the first occlusion feature map and the first human body feature map obtained by feature decoupling, semantic segmentation is performed respectively. Based on the second occlusion feature map and the second human body feature map obtained by semantic segmentation, key point detection is performed to determine the key point detection result.

[0120] The device embodiments described above are merely illustrative. 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 modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0121] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0122] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting key points on the human body, characterized in that, include: Acquire an image of the human body to be detected, wherein the image of the human body to be detected is an image of a target human body in an occluded scene; The human image to be detected is input into a preset key point detection network model to obtain the key point detection results of the target human body output by the preset key point detection network model; The preset key point detection network model is used to extract different channel feature maps in the human image to be detected, and to decouple the different channel feature maps. Based on the first occlusion feature map and the first human feature map obtained by the feature decoupling, semantic segmentation is performed respectively. Based on the second occlusion feature map and the second human feature map obtained by the semantic segmentation, key point detection is performed to determine the key point detection result. The step of extracting different channel feature maps from the human image to be detected includes: Based on a preset spatial regularization loss function, different local spatial locations in the human body image to be detected are mapped to different channels to determine different channel feature maps in the human body image to be detected; the preset spatial regularization loss function is used to limit the uncorrelation between convolutional kernels and the uncorrelation between channels; The feature decoupling of the different channel feature maps includes: The different channel feature maps are weighted based on different preset occlusion channel weights; The different channel feature maps are weighted based on different preset human body channel weights; The semantic segmentation of the first occlusion feature map and the first human body feature map obtained based on the feature decoupling includes: The first occluder feature map and the first human body feature map obtained by decoupling the features are downsampled to determine the target occluder feature map and the target human body feature map. Semantic segmentation is performed on the feature map of the target occluder and the feature map of the target human body, respectively.

2. The method for detecting key human body points according to claim 1, characterized in that, The training process of the preset key point detection network model includes: A sample human body image and an initial keypoint detection network model are determined. The initial keypoint detection network model includes a feature extraction model, a feature decoupling model, an occlusion semantic segmentation model, a human body semantic segmentation model, and a keypoint detection model. The sample human body image is input into the initial keypoint detection network model for training of feature extraction, feature decoupling, downsampling, semantic segmentation, and keypoint detection, thereby determining the preset keypoint detection network model; wherein, the feature decoupling includes weighting the different sample channel feature maps obtained by feature extraction using different sample occlusion channel weights and different sample human body channel weights respectively.

3. The method for detecting key human body points according to claim 2, characterized in that, The step of inputting the sample human image into the initial keypoint detection network model for feature extraction, feature decoupling, downsampling, semantic segmentation, and keypoint detection training, and determining the preset keypoint detection network model, includes: The sample human body image is input into the initial key point detection network model for iterative training of feature extraction, feature decoupling, downsampling, semantic segmentation and key point detection. The sample occlusion segmentation result and sample human body segmentation result obtained after training for a preset number of iterations, as well as the key point detection result of the intermediate key point detection network model obtained after training for a preset number of iterations, are obtained. Determine whether the occlusion segmentation index value of the sample occlusion segmentation result reaches the preset occlusion segmentation index threshold, whether the human body segmentation index value of the sample human body segmentation result reaches the preset human body segmentation index threshold, and whether the detection index value of the key point detection result reaches the preset detection threshold. When the occlusion segmentation index value reaches the preset occlusion segmentation index threshold, the human body segmentation index value reaches the preset human body segmentation index threshold, and the detection index value reaches the preset detection threshold, training is stopped, and the intermediate key point detection network model corresponding to the training stop is determined as the preset key point detection network model. When it is determined that the detection index value does not reach the preset detection threshold, the occlusion segmentation index value does not reach the preset occlusion segmentation index threshold, and / or the human body segmentation index value does not reach the preset human body segmentation index threshold, the parameters of the intermediate key point detection network model are updated, and the sample occlusion channel weights and / or sample human body channel weights participating in the training for the preset number of iterations are adjusted. Based on the adjusted sample occlusion channel weights, the adjusted sample human body channel weights, and / or the updated model parameters, the intermediate key point detection network model is iterated and trained again; until training stops, the preset key point detection network model is determined.

4. The method for detecting key human body points according to claim 3, characterized in that, The feature decoupling model includes an occlusion channel attention model and a human body channel attention model. The iterative training of inputting the sample human body image into the initial keypoint detection network model for feature extraction, feature decoupling, downsampling, semantic segmentation, and keypoint detection includes: For each training iteration, the sample human body image is input into the feature extraction model of the initial keypoint detection network model to obtain different sample channel feature maps in the sample human body image; The different sample channel feature maps are input into the occlusion channel attention model for convolution, downsampling and full connection to determine the different sample occlusion channel weights; The different sample channel feature maps are input into the human body channel attention model for convolution, downsampling and full connection to determine the different sample human body channel weights. The different sample occlusion channel weights and the different sample human body channel weights are used to weight the different sample channel feature maps respectively to determine the first sample occlusion feature map and the first sample human body feature map. The first occluder feature map and the first human body feature map of the sample are downsampled respectively to determine the target occluder feature map and the target human body feature map of the sample. Semantic segmentation is performed on the sample target occlusion feature map and the sample target human body feature map to determine the sample second occlusion feature map and the sample second human body feature map; Key point detection is performed on the second occlusion feature map and the second human body feature map of the sample.

5. The method for detecting key human body points according to claim 3 or 4, characterized in that, The method further includes: While determining that the intermediate key point detection network model corresponding to the training stop is the preset key point detection network model, the adjusted sample occlusion channel weights corresponding to the training stop are determined to be different preset occlusion channel weights, and the adjusted sample human body channel weights corresponding to the training stop are determined to be different preset human body channel weights.

6. A human body key point detection device, characterized in that, include: The acquisition module is used to acquire the human body image to be detected, which is an image of the target human body in an occluded scene; The detection module is used to input the human image to be detected into a preset key point detection network model to obtain the key point detection results of the target human body output by the preset key point detection network model; The preset keypoint detection network model is used to extract different channel feature maps from the human image to be detected, and to decouple the different channel feature maps. Semantic segmentation is performed on the first occlusion feature map and the first human feature map obtained from the feature decoupling. Keypoint detection is performed on the second occlusion feature map and the second human feature map obtained from the semantic segmentation, and the keypoint detection result is determined. Extracting different channel feature maps from the human image to be detected includes: mapping different local spatial locations in the human image to be detected to different channels based on a preset spatial regularization loss function, thereby determining the different channel feature maps in the human image to be detected. The regularization loss function is used to limit the correlation between convolutional kernels and the correlation between channels; the feature decoupling of the different channel feature maps includes: weighting the different channel feature maps based on different preset occlusion channel weights; weighting the different channel feature maps based on different preset human body channel weights; the semantic segmentation of the first occlusion feature map and the first human body feature map obtained by the feature decoupling includes: downsampling the first occlusion feature map and the first human body feature map obtained by the feature decoupling to determine the target occlusion feature map and the target human body feature map; and performing semantic segmentation on the target occlusion feature map and the target human body feature map respectively.

7. An electronic 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 human body key point detection method as described in any one of claims 1 to 5.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the human body key point detection method as described in any one of claims 1 to 5.

Images