An AI-based intelligent human posture recognition system

Abstract

This invention discloses an artificial intelligence-based intelligent human posture recognition system, comprising a data collection module, an image data optimization module, a human posture key point extraction module, and a human posture intelligent recognition module. This invention relates to the technical field of human posture recognition, specifically to an artificial intelligence-based intelligent human posture recognition system. This solution innovatively designs a multi-scale feature enhancement block for human joints. Through multi-scale convolutional layers, combined with joint channel feature calibration and pixel feature calibration, it improves the multi-scale joint extraction capability, effectively handling posture changes in complex dynamic scenes. It also designs a joint angular velocity dynamic attention layer and a hierarchical spatiotemporal graph convolution. By dynamically adjusting the feature channel weights based on human joint angular velocity and capturing the spatial linkage and temporal changes of the target human body's movements layer by layer, it significantly improves the modeling accuracy of temporal features and the adaptability to dynamic scenes, effectively enhancing the accuracy of human posture recognition.

Description

Technical Field

[0001] This invention relates to the technical field of human posture recognition, specifically to an artificial intelligence-based intelligent human posture recognition system. Background Technology

[0002] An AI-based intelligent human posture recognition system utilizes artificial intelligence technology to identify and analyze athletes' human postures in real time during training or competition. It collects athletes' posture data in real time via cameras, combines this data with an AI model to extract and identify key points of the human posture, and performs high-precision human posture and movement analysis. The system accurately captures changes in the athlete's body at different stages of movement, achieving comprehensive recognition of human postures. This provides precise data support for athletes' training and improves the training efficiency and management quality for coaches, athletes, and related managers.

[0003] However, traditional human posture key point extraction suffers from technical problems such as insufficient joint feature extraction capability and lack of global limb association feature capture, which makes it impossible to accurately extract human posture key points in complex dynamic scenes, thus affecting the accuracy of posture recognition. Existing human posture recognition models rely only on static features or single-modal features, which cannot effectively capture the spatial linkage and temporal evolution of human joints in complex scenes, resulting in low accuracy of human posture recognition results. Summary of the Invention

[0004] To address the aforementioned issues and overcome the shortcomings of existing technologies, this invention provides an AI-based intelligent human posture recognition system. It addresses the technical problems of insufficient joint feature extraction and lack of global limb correlation feature capture in traditional human posture key point extraction methods, which lead to inaccurate extraction of human posture key points in complex dynamic scenes and thus affect the accuracy of posture recognition. This solution innovatively designs a multi-scale feature enhancement block for human joints, performing parallel multi-scale joint feature extraction through parallel multi-scale convolutional layers, and combining joint channel feature calibration and joint pixel feature calibration. This effectively avoids small joint omissions and large joint positioning errors, improves multi-scale joint extraction capabilities, enhances joint feature recognition capabilities, and improves the ability to capture global limb correlation features, effectively handling complex dynamic scenes. This solution addresses the issue of low accuracy in human posture recognition due to the inability of existing human posture recognition models, which rely solely on static or single-modal features and thus fail to effectively capture the spatial linkages and temporal evolution of human joints in complex scenes. Instead, it innovatively designs a joint angular velocity dynamic attention layer and hierarchical spatiotemporal graph convolution. By dynamically adjusting feature channel weights based on human joint angular velocity and hierarchically capturing the spatial linkages and temporal changes of target human movements, it enhances the ability to model temporal action sequences, significantly improving the modeling accuracy of temporal features and the adaptability to dynamic scenes. This effectively improves the accuracy of human posture recognition, enabling precise identification of target human postures in high-speed motion and complex action scenarios.

[0005] The technical solution adopted by the present invention is as follows: The present invention provides an artificial intelligence-based human posture intelligent recognition system, including a data collection module, an image data optimization module, a human posture key point extraction module, and a human posture intelligent recognition module;

[0006] The data collection module specifically obtains raw human posture recognition data by collecting data.

[0007] The image data optimization module specifically obtains human posture optimized image data through image data cleaning, image deduplication, image denoising, and image normalization.

[0008] The human posture key point extraction module specifically involves constructing a human posture key point extraction model through shallow feature initialization extraction, human joint multi-scale feature calibration, human whole body limb association feature capture, human joint fusion feature extraction, and human posture key point output, and then training the model. Finally, the target posture recognition data is input into the trained model to obtain the target human posture key point set.

[0009] The human posture intelligent recognition module specifically establishes a human posture recognition model by constructing a human motion space graph convolutional layer, a joint angular velocity dynamic attention layer, a human posture temporal convolutional layer, and a human posture classification output layer, and trains the recognition model. Finally, the set of key points of the target human posture is input into the trained human posture recognition model to obtain the target human posture intelligent recognition result.

[0010] Furthermore, the data collection module specifically obtains raw human posture recognition data through data collection operations; the raw human posture recognition data includes reference posture recognition data and target posture recognition data; both the reference posture recognition data and the target posture recognition data include human posture image data, and the reference posture recognition data also includes reference human joint key point annotation data and reference posture annotation data.

[0011] Furthermore, the image data optimization module includes image data cleaning, image deduplication, image denoising, and image normalization to obtain human pose optimized image data;

[0012] The image data cleaning specifically involves removing incomplete or low-quality images from human posture image data using an image quality detection algorithm.

[0013] The image deduplication specifically involves comparing and filtering consecutive or similar image frames using a similarity detection algorithm to delete duplicate images.

[0014] The image denoising specifically involves smoothing the image using Gaussian filtering technology;

[0015] The image normalization specifically involves adjusting the range of pixel values ​​in the image to a preset standard range through standardization processing.

[0016] Furthermore, the human posture key point extraction module specifically includes the following steps:

[0017] Shallow feature initialization extraction specifically involves performing initial feature mapping on the image through convolution operations combined with the ReLU activation function to obtain a shallow feature map of human pose.

[0018] Human joint multi-scale feature calibration specifically involves inputting a shallow human posture feature map into eight cascaded human joint multi-scale feature enhancement blocks to obtain a human joint multi-scale enhanced feature map; the human joint multi-scale feature enhancement blocks include the following steps:

[0019] The multi-scale parallel feature extraction of human joints is specifically carried out by using G parallel convolutional layers to extract multi-scale parallel features of human joints, and performing a concatenation operation on the output of G layers to obtain a multi-scale fused feature map.

[0020] Joint channel feature calibration specifically involves performing global average pooling and global max pooling operations on the multi-scale fused feature map, and inputting the pooling results into a fully connected layer and a ReLU activation function to obtain two channel feature vectors. After concatenating the two channel feature vectors, joint channel attention weights are generated through a fully connected layer and a Sigmoid activation function. The channel attention weights are then multiplied element-wise with the multi-scale fused feature map to obtain the joint channel feature map.

[0021] Joint pixel feature calibration specifically involves performing a convolution operation on the joint channel feature map, generating pixel attention weights by passing the convolution result through the Sigmoid activation function, and performing element-wise multiplication operations between the pixel attention weights and the joint channel feature map to obtain the joint pixel calibration feature map.

[0022] Residual connection feature fusion specifically involves concatenating the joint pixel calibration feature map with the human joint multi-scale enhancement feature map output by the previous human joint multi-scale feature enhancement block to obtain the human joint multi-scale enhancement feature map of the current human joint multi-scale feature enhancement block.

[0023] The capture of human body limb association features involves performing convolution operations on the multi-scale enhanced feature maps of human joints through four parallel dilated convolutional layers, and then performing a concatenation operation on the output features of the four dilated convolutional layers to obtain the human body whole-body association feature map.

[0024] Human joint fusion feature extraction specifically involves performing a convolution operation on the shallow feature map of human posture and performing an element-wise multiplication operation with the associated feature map of the whole human body to obtain the human joint fusion feature map.

[0025] The key points of human pose are output by performing a convolution operation on the human joint fusion feature map, and then mapping the output values ​​to a sigmoid activation function. In the interval, generate a confidence map of human key points, and for the channel corresponding to the k-th joint, traverse all pixels in the channel and select the pixel position with the highest confidence as the coordinate of the joint. Integrate the coordinates of N human joints with their corresponding confidence to obtain a set of human pose key points, where k represents the index of the joint.

[0026] The model is constructed and trained by means of shallow feature initialization extraction, human joint multi-scale feature calibration, human body limb association feature capture, human joint fusion feature extraction and human posture key point output to construct a human posture key point extraction model. The reference posture recognition data based on the human posture optimized image data is used as training data to train the model and obtain the trained human posture key point extraction model.

[0027] The extraction of key points of target human posture involves inputting the target posture recognition data of the optimized human posture image data into the trained human posture key point extraction model to obtain a set of key points of target human posture.

[0028] Furthermore, the human posture intelligent recognition module specifically includes the following steps:

[0029] Establishing a human pose recognition model includes the following steps:

[0030] Constructing a convolutional layer for the human motion space graph includes the following steps:

[0031] The spatiotemporal graph of human posture is established by constructing graph nodes, constructing spatial relation edges, and constructing a spatial adjacency matrix, resulting in a set of graph nodes, a set of spatial relation edges, and a set of normalized spatial adjacency matrices.

[0032] The construction of the spatial adjacency matrix specifically involves constructing an initial adjacency matrix based on the spatial relation edge set, and dividing the initial adjacency matrix into a centripetal joint submatrix, an eccentric joint submatrix, and a joint self-connection submatrix according to the straight-line distance from the joint to the center of gravity of the human body. Then, normalization processing is performed on the submatrixes respectively to obtain a set of normalized spatial adjacency matrices.

[0033] Spatial graph convolution feature extraction specifically involves inputting the graph node set and the normalized spatial adjacency matrix set into a graph convolutional layer constructed based on a graph convolutional network. By configuring independent learnable edge masks and node feature transformation weights for different submatrices, multi-scale spatial graph convolution operations are performed. Finally, a nonlinear transformation is introduced by combining the ReLU activation function to obtain spatial graph convolution features.

[0034] Constructing a joint angular velocity dynamic attention layer; specifically including the following steps:

[0035] The average joint angular velocity over time is calculated by using the joint coordinates of consecutive frames to calculate the angular velocity of the k-th joint in frame t, and then calculating the average joint angular velocity of the k-th joint over the action cycle of frame T, where T represents the preset total number of frames in the action cycle, and t represents the frame index.

[0036] To calculate the channel response intensity, specifically, global average pooling is performed on the spatial graph convolutional features according to the joint and temporal dimensions to obtain the channel response intensity.

[0037] The dynamic weight generation process involves first calculating the kernel size of the one-dimensional convolution based on the number of channels in the spatial graph convolution feature, then modulating the channel response intensity using the joint angular velocity time mean, finally capturing cross-channel interaction relationships through one-dimensional convolution, and combining the weights with the Sigmoid activation function to normalize the weights, thus obtaining dynamic cross-channel weights.

[0038] Feature calibration specifically involves performing element-wise multiplication of dynamic cross-channel weights with spatial graph convolutional features, scaling each channel feature according to the weights, and obtaining enhanced spatial correlation features.

[0039] To construct a temporal convolutional layer for human pose, the enhanced spatial correlation features are concatenated into a temporal feature sequence according to the time dimension, and a one-dimensional convolution is used to perform the temporal convolution operation. The ReLU activation function is combined to enhance the nonlinear expression and obtain the temporal features of human pose.

[0040] A human pose classification output layer is constructed. Specifically, global average pooling is performed on the temporal features of human pose, and feature dimension transformation is performed through two fully connected layers. Finally, the probability of each category is output through the Softmax activation function, and the category with the highest probability is selected as the human pose output result.

[0041] The recognition model training specifically involves using the reference posture recognition data in the human posture optimization image data as training data to train the recognition model and obtain the trained human posture recognition model.

[0042] Intelligent recognition of target human posture involves inputting the set of key points of the target human posture into a trained human posture recognition model to obtain the intelligent recognition result of the target human posture, thereby realizing intelligent recognition of the posture of the target human body.

[0043] The beneficial effects achieved by the present invention using the above solution are as follows:

[0044] (1) In view of the technical problems of insufficient joint feature extraction capability and lack of global limb association feature capture in traditional human posture key point extraction, which makes it impossible to accurately extract human posture key points in complex dynamic scenes, thus affecting the accuracy of posture recognition, this solution innovatively designs a multi-scale feature enhancement block for human joints. It performs multi-scale parallel feature extraction of joints through parallel multi-scale convolutional layers, and combines joint channel feature calibration and joint pixel feature calibration to effectively avoid small joint omissions and large joint positioning deviations, improve the multi-scale joint extraction capability, enhance the joint feature recognition capability, and improve the global association feature capture capability of limbs. It can effectively cope with posture changes in complex dynamic scenes, thereby accurately extracting key points during movement and significantly improving the accuracy of human posture recognition.

[0045] (2) In view of the technical problem that existing human posture recognition models rely only on static features or single-modal features, which cannot effectively capture the spatial linkage and temporal evolution of human joints in complex scenes, resulting in low accuracy of human posture recognition results, this solution innovatively designs a joint angular velocity dynamic attention layer and a hierarchical spatiotemporal graph convolution. By dynamically adjusting the feature channel weights based on human joint angular velocity and capturing the spatial linkage and temporal changes of target human movements in layers, the action temporal modeling capability is strengthened, the modeling accuracy of temporal features and the adaptability of dynamic scenes are significantly improved, the accuracy of human posture recognition is effectively improved, and the target human posture is accurately recognized in high-speed motion and complex action scenes. Attached Figure Description

[0046] Figure 1 A schematic diagram of a human posture intelligent recognition system based on artificial intelligence provided by the present invention;

[0047] Figure 2 A flowchart illustrating the process of extracting key points of human posture.

[0048] Figure 3 A flowchart illustrating the process of calibrating multi-scale features of human joints in the human posture key point extraction module.

[0049] Figure 4 This is a flowchart illustrating the human posture intelligent recognition module.

[0050] Figure 5 A flowchart illustrating the process of establishing a human posture recognition model in the human posture intelligent recognition module;

[0051] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0052] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0053] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the system or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0054] Example 1, see Figure 1 The technical solution adopted by the present invention is as follows: The present invention provides an artificial intelligence-based human posture intelligent recognition system, including a data collection module, an image data optimization module, a human posture key point extraction module and a human posture intelligent recognition module;

[0055] The data collection module is used to acquire the basic raw data required for human posture recognition. Specifically, it collects data to obtain raw data for human posture recognition and sends the data to the image data optimization module.

[0056] The image data optimization module receives data sent by the data collection module and is used to preprocess single-frame images in the human posture image data to improve image quality. Specifically, it obtains human posture optimized image data through image data cleaning, image deduplication, image denoising and image normalization, and sends the data to the human posture key point extraction module and the human posture intelligent recognition module.

[0057] The human posture key point extraction module receives data sent by the image data optimization module and is used to extract accurate human posture joint key points from the human posture optimization image data. Specifically, it constructs a human posture key point extraction model through shallow feature initialization extraction, human joint multi-scale feature calibration, human whole body limb association feature capture, human joint fusion feature extraction, and human posture key point output, and performs model training. Finally, the target posture recognition data is input into the trained model to obtain the target human posture key point set, and the data is sent to the human posture intelligent recognition module.

[0058] The human posture intelligent recognition module receives data sent by the image data optimization module and the human posture key point extraction module, and uses it to establish a human posture recognition model and achieve accurate recognition of the target human posture. Specifically, it establishes a human posture recognition model by constructing a human motion space graph convolutional layer, a joint angular velocity dynamic attention layer, a human posture temporal convolutional layer, and a human posture classification output layer, and trains the recognition model. Finally, it inputs the set of key points of the target human posture into the trained human posture recognition model to obtain the intelligent recognition result of the target human posture.

[0059] Example 2, see Figure 1This embodiment is based on the above embodiment. Specifically, the data collection module collects data through video surveillance equipment and uses video parsing and frame extraction algorithms to obtain raw human posture recognition data according to a preset sampling frequency and time interval. The raw human posture recognition data includes reference posture recognition data and target posture recognition data. Both the reference posture recognition data and the target posture recognition data include human posture image data. The reference posture recognition data also includes reference human joint key point annotation data and reference posture annotation data.

[0060] The human posture image data refers to single-frame image data extracted from videos of athletes or monitored objects during training, competitions or other activities collected by video monitoring equipment. It includes the athlete's posture, movement changes and the movement trajectory of each joint in dynamic scenes, and is the basic data source for human posture analysis.

[0061] The reference human joint key point annotation data refers to the specific coordinate information of the position of each joint in the human body.

[0062] The reference posture annotation data refers to the labeled data of the posture of the monitored object, which describes the posture category of the target human body at a specific moment.

[0063] Example 3, see Figure 1 This embodiment is based on the above embodiment. The image data optimization module is used to preprocess single-frame images in human posture image data to improve image quality, including image data cleaning, image deduplication, image denoising and image normalization, to obtain human posture optimized image data.

[0064] The image data cleaning is used to clean single-frame images in human posture image data. Specifically, it removes incomplete or low-quality images from human posture image data through an image quality detection algorithm.

[0065] The image deduplication is used to remove duplicate image frames from human pose image data. Specifically, it uses a similarity detection algorithm to compare and filter consecutive or similar image frames to delete duplicate images.

[0066] The image denoising is used to remove noise from human posture image data, specifically by smoothing the image using Gaussian filtering technology.

[0067] The image normalization is used to adjust the pixel values ​​of human posture images to a uniform range. Specifically, it involves adjusting the range of pixel values ​​of the image to a preset standard range through standardization processing.

[0068] Example 4, see Figure 1 , Figure 2 and Figure 3This embodiment is based on the above embodiment, and the human posture key point extraction module specifically includes the following steps:

[0069] Shallow feature initialization extraction is used to perform initial feature transformation on the human pose optimization image. While preserving the spatial information of key regions of the human pose, redundant channel features are compressed to reduce the interference of background noise on subsequent keypoint localization. Specifically, convolution operations combined with the ReLU activation function are used to perform initial feature mapping on the image, obtaining a shallow feature map of the human pose. The formula used is as follows:

[0070] ;

[0071] In the formula, This represents a shallow feature map of human posture, with dimensions of [dimension number missing]. It includes basic features of the human body outline and clues to the initial position of joints. This represents the ReLU activation function, used to introduce nonlinear feature transformations and enhance the ability to express the nonlinear features of human posture. This represents a 1×1 convolution operation with 64 kernels and a stride of 1. This represents human pose optimization image data, where W is the image width, H is the image height, and 3 is the number of RGB channels.

[0072] Human joint multi-scale feature calibration is used to extract and optimize human joint features in the shallow feature map of human posture at multiple scales, highlighting the channel and pixel features of the joint region, suppressing background noise, and improving the recognizability of joint features. Specifically, the shallow feature map of human posture is input into eight cascaded human joint multi-scale feature enhancement blocks to obtain a human joint multi-scale enhanced feature map. The human joint multi-scale feature enhancement blocks include the following steps:

[0073] Multi-scale parallel feature extraction of human joints is used to capture joint features at different scales in the human body. Specifically, G parallel convolutional layers are used to extract multi-scale parallel features of human joints, and the outputs of the G layers are concatenated to obtain a multi-scale fused feature map. The formula used is as follows:

[0074] ;

[0075] In the formula, This represents a multi-scale fused feature map. This indicates a feature concatenation operation. This represents the output features of a 1×1 convolutional layer, with output channels. It is used to capture the general outline features of the human torso. This represents the output features of a 3×3 convolutional layer with 2 output channels. The convolutional layers are used to capture details of small joints in the human body. This represents the output features of a 3×3 convolutional layer, with the following output channels: Output channel The convolutional layers are used to capture the contours of large human joints, and G represents the preset number of parallel convolutional layers.

[0076] Joint channel feature calibration is used to enhance the channel-level feature response of the joint region and suppress invalid interference from the background channel. Specifically, global average pooling and global max pooling operations are performed on the multi-scale fused feature map, and the pooling results are input into a fully connected layer and a ReLU activation function, respectively, to obtain two channel feature vectors. After concatenating the two channel feature vectors, joint channel attention weights are generated through a fully connected layer and a sigmoid activation function. The channel attention weights are then multiplied element-wise with the multi-scale fused feature map to obtain the joint channel feature map. The formula used is as follows:

[0077] ;

[0078] ;

[0079] In the formula, Indicates the attention weight of the joint channel. This represents the Sigmoid activation function. This indicates a fully connected layer operation. This represents a global average pooling operation, used to capture global channel dependencies of human joints. This represents a global max-pooling operation, used to enhance the channel response of human joints. Represents the joint channel feature map. This represents an element-wise multiplication operation;

[0080] Joint pixel feature calibration is used to enhance the pixel-level feature response of the joint region and suppress invalid interference from background pixels. Specifically, a 1×1 convolution operation is performed on the joint channel feature map, and the convolution result is used to generate pixel attention weights through a sigmoid activation function. The pixel attention weights are then multiplied element-wise with the joint channel feature map to obtain the joint pixel calibration feature map. The formula used is as follows:

[0081] ;

[0082] ;

[0083] In the formula, Represents pixel attention weights. This represents the joint pixel calibration feature map. This represents a 1×1 convolution operation;

[0084] Residual connection feature fusion is used to preserve historical feature information through residual connections, avoiding feature degradation caused by multiple rounds of calibration. Specifically, it concatenates the joint pixel calibration feature map with the human joint multi-scale enhancement feature map output by the previous human joint multi-scale feature enhancement block to obtain the human joint multi-scale enhancement feature map of the current human joint multi-scale feature enhancement block; the formula used is as follows:

[0085] ;

[0086] In the formula, This represents the multi-scale enhancement feature map of the human joint in the i-th enhancement block. Indicates the first Multi-scale enhancement feature map of human joints in an enhanced block This represents the joint pixel calibration feature map of the i-th enhancement block. This indicates a residual splicing operation;

[0087] The human body full-body limb association feature capture method is used to expand the effective receptive field, capture the linkage features of the upper limbs, trunk, and lower limbs, and enhance the association expression of the whole body posture. Specifically, it performs convolution operations on the multi-scale enhanced feature maps of human joints through four parallel 3×3 dilated convolutional layers, and performs a concatenation operation on the output features of the four dilated convolutional layers to obtain the human body full-body association feature map; the formula used is as follows:

[0088] ;

[0089] In the formula, Represents a map of interconnected features of the entire human body. This represents a 3×3 dilated convolution operation with a dilation rate of 1. This represents a 3×3 dilated convolution operation with a dilation rate of 3. This represents a 3×3 dilated convolution operation with a dilation rate of 5. This indicates a 3×3 dilated convolution operation with a dilation rate of 7. This represents a multi-scale enhancement feature map of the human joint in the 8th enhancement block.

[0090] Human joint fusion feature extraction is used to fuse shallow human posture features with whole-body associated features. Specifically, a 1×1 convolution operation is performed on the shallow human posture feature map, and then an element-wise multiplication operation is performed with the whole-body associated feature map to obtain the human joint fusion feature map. The formula used is as follows:

[0091] ;

[0092] In the formula, This represents a human joint fusion feature map, including initial joint cues and systemic association features. This represents a 1×1 convolution operation, used to... The number of channels was adjusted to match Channel matching;

[0093] The human pose keypoint output is used to extract human joint keypoints. Specifically, a 3×3 convolution operation is performed on the human joint fusion feature map to map the feature dimensions to a preset number of human joints, and then the output values ​​are mapped to a sigmoid activation function. The algorithm generates a confidence map of human body key points within a given interval. For the channel corresponding to the k-th joint, it iterates through all pixels in that channel and selects the pixel position with the highest confidence as the coordinate of that joint. The coordinates of N human body joints are integrated with their corresponding confidence scores to obtain a set of human posture key points. The human body key points include the head, left shoulder joint, right shoulder joint, chest, abdomen, buttocks, left upper arm, right upper arm, left lower arm, right lower arm, left hand, right hand, left upper leg, right upper leg, left lower leg, and right lower leg.

[0094] The formula used is as follows:

[0095] ;

[0096] ;

[0097] ;

[0098] In the formula, This indicates a 3×3 convolution operation, which maps the feature dimensions to a preset number N of human body joints. This represents a confidence map of human key points, reflecting the confidence level of each pixel as a corresponding joint. Each channel corresponds to one preset joint. This represents the pixel coordinates with the highest built-in confidence level in the k-th joint channel. This indicates the confidence value corresponding to the coordinate with the highest confidence level. Represents the coordinates of the k-th joint. It is the position with the highest built-in confidence level in its channel, corresponding to a confidence level of . k represents the joint index. Represents the set of key points in human posture;

[0099] The model is constructed and trained by means of shallow feature initialization extraction, human joint multi-scale feature calibration, human body limb association feature capture, human joint fusion feature extraction and human posture key point output to construct a human posture key point extraction model. The model is trained by using the reference posture recognition data of the human posture optimized image data as training data to obtain the trained human posture key point extraction model.

[0100] The model training specifically employs a gradient descent-type optimizer, using the weighted sum of the mean square error of keypoint coordinates and the cross-entropy of confidence as the total loss function, iteratively optimizing the model parameters, and gradually updating the model until training is complete.

[0101] The extraction of key points of target human posture involves inputting the target posture recognition data of the optimized human posture image data into the trained human posture key point extraction model to obtain a set of key points of target human posture.

[0102] By performing the above operations, this solution addresses the technical problems of insufficient joint feature extraction and lack of global limb correlation feature capture in traditional human posture key point extraction, which leads to the inability to accurately extract human posture key points in complex dynamic scenes, thus affecting the accuracy of posture recognition. This solution innovatively designs a multi-scale feature enhancement block for human joints, and performs multi-scale parallel feature extraction of joints through parallel multi-scale convolutional layers. Combined with joint channel feature calibration and joint pixel feature calibration, it effectively avoids small joint omissions and large joint positioning deviations, improves multi-scale joint extraction capabilities, enhances joint feature recognition capabilities, and improves the ability to capture global limb correlation features. It can effectively cope with posture changes in complex dynamic scenes, thereby accurately extracting key points during movement and significantly improving the accuracy of human posture recognition.

[0103] Example 5, see Figure 1 , Figure 4 and Figure 5 This embodiment is based on the above embodiment, and the human posture intelligent recognition module specifically includes the following steps:

[0104] A human pose recognition model is established to construct a multi-layered neural network model for human pose recognition. This includes the following steps:

[0105] Constructing a convolutional layer for the human motion space graph includes the following steps:

[0106] The human body posture spatiotemporal graph is established to convert discrete key points of the athlete's human posture into spatiotemporal graph structure data. Specifically, it is achieved by constructing graph nodes, constructing spatial relation edges, and constructing a spatial adjacency matrix to obtain a graph node set, a spatial relation edge set, and a normalized spatial adjacency matrix set.

[0107] The construction of graph nodes specifically involves using key points from the set of key points of the target human body posture as graph nodes to obtain a graph node set;

[0108] The construction of spatial relationship edges specifically involves generating a set of spatial relationship edges by pairing key points in the graph node set based on the natural connection rules of human anatomy.

[0109] The construction of the spatial adjacency matrix specifically involves constructing an initial adjacency matrix based on the spatial relation edge set, and dividing the initial adjacency matrix into a centripetal joint submatrix, an eccentric joint submatrix, and a joint self-connection submatrix according to the straight-line distance from the joint to the center of gravity of the human body. Then, normalization processing is performed on the submatrixes respectively to obtain a set of normalized spatial adjacency matrices.

[0110] The elements in the initial adjacency matrix This indicates that joints i and j have a connecting edge; otherwise, the value is 0.

[0111] The centripetal joint submatrix Specifically, the connection relationships of joints whose straight-line distance from the human body's center of gravity is less than or equal to 0.3 times the human body's height are divided into centripetal joint submatrices;

[0112] The centrifugal joint sub-matrix Specifically, the connection relationships of joints whose straight-line distance from the center of gravity to the human body is greater than 0.3 times the height of the human body are divided into eccentric joint sub-matrices;

[0113] The joint self-connection submatrix Specifically, the diagonal elements are set to 1, and the matrix is ​​divided into joint self-connection submatrices;

[0114] Spatial graph convolution feature extraction is used to capture the spatial correlation features of joints in the dynamic movements of athletes. Specifically, the graph node set and the normalized spatial adjacency matrix set are input into a graph convolutional layer constructed based on a graph convolutional network. By configuring independent learnable edge masks and node feature transformation weights for different sub-matrices, multi-scale spatial graph convolution operations are performed. Finally, a nonlinear transformation is introduced by combining the ReLU activation function to obtain the spatial graph convolution features. The formula used is as follows:

[0115] ;

[0116] In the formula, Represents spatial graph convolution features. This represents the j-th submatrix after normalization. This represents the learnable edge mask of the j-th submatrix. Represents a graph node set. Represents the node feature transformation weights of the j-th submatrix;

[0117] A dynamic attention layer for joint angular velocity is constructed to dynamically adjust the feature channel weights based on the human joint angular velocity, thereby enhancing the effective features of high-speed eccentric joints. This includes the following steps:

[0118] The time-averaged joint angular velocity is calculated to quantify the average dynamic intensity of the k-th joint in the human body during a complete motion cycle, providing a global reference for adaptive modulation of channel weights. Specifically, the angular velocity of the k-th joint in frame t is calculated using joint coordinates from consecutive frames, and the average joint angular velocity of the k-th joint during the motion cycle in frame T is also calculated. The formula used is as follows:

[0119] ;

[0120] ;

[0121] In the formula, This represents the angular velocity of the k-th joint in frame t. Let represent the linear velocity vector of the k-th joint in frame t. Let represent the acceleration vector of the k-th joint in frame t. Let represent the average joint angular velocity of the k-th joint during a complete motion cycle, and T represent the total number of frames contained in the complete motion cycle. Indicates the starting frame number of the action cycle;

[0122] The channel response intensity is calculated to extract channel-level global feature information from the spatial graph convolutional features. Specifically, global average pooling is performed on the spatial graph convolutional features according to the joint and temporal dimensions to obtain the channel response intensity; the formula used is as follows:

[0123] ;

[0124] In the formula, This represents the response intensity of the g-th channel. This represents the feature value of the g-th channel in the temporal sequence of the k-th joint in the t-th frame of the spatial graph convolution feature;

[0125] Dynamic weight generation is used to generate cross-channel interaction weights that fit dynamic actions. Specifically, it first calculates the kernel size of a one-dimensional convolution based on the number of channels in the spatial graph convolution feature, then modulates the channel response intensity using the joint angular velocity time mean, and finally captures the cross-channel interaction relationship through a one-dimensional convolution. The weights are then normalized using a sigmoid activation function to obtain the dynamic cross-channel weights. The formula used is as follows:

[0126] ;

[0127] ;

[0128] ;

[0129] In the formula, This represents the size of the one-dimensional convolution kernel. The number of channels representing the convolutional features of the spatial graph. This represents the channel number adjustment coefficient, with a value of 2. This represents the offset coefficient, with a value of 1. This indicates rounding down to the nearest odd number. This represents the modulation response intensity of the g-th channel. This represents the time-averaged modulation coefficient of the joint angular velocity, with a value of 0.8. This represents the dynamic cross-channel weight of the g-th channel. Indicates the modulated first The response intensity of each channel, where i represents the channel offset. This represents the one-dimensional convolution parameter corresponding to the g-th channel;

[0130] Feature calibration, used to optimize spatial graph convolutional features and highlight key features of high-speed joints in athletes, specifically involves element-wise multiplication of dynamic cross-channel weights with spatial graph convolutional features, scaling each channel feature according to the weights to obtain enhanced spatial correlation features; the formula used is as follows:

[0131] ;

[0132] In the formula, This indicates enhanced spatial correlation features;

[0133] A temporal convolutional layer for human pose is constructed to capture the temporal evolution features of athlete movements. Specifically, the enhanced spatial correlation features are concatenated into a temporal feature sequence along the time dimension, and a one-dimensional convolution operation is performed to perform the temporal convolution operation. The ReLU activation function is then used to enhance the nonlinear expression, resulting in the temporal features of human pose. The formulas used are as follows:

[0134] ;

[0135] In the formula, Represents a time-series characteristic sequence. Represents the temporal characteristics of human posture. This represents a one-dimensional temporal convolution operation, where d represents the dilation rate, and the kernel size is set to 2d+1.

[0136] A human pose classification output layer is constructed. Specifically, global average pooling is performed on the temporal features of human pose, and feature dimension transformation is performed through two fully connected layers. Finally, the probability of each category is output through the Softmax activation function, and the category with the highest probability is selected as the human pose output result.

[0137] The recognition model training specifically involves using the reference posture recognition data in the human posture optimization image data as training data to train the recognition model and obtain the trained human posture recognition model.

[0138] Intelligent recognition of target human posture involves inputting the set of key points of the target human posture into a trained human posture recognition model to obtain the intelligent recognition result of the target human posture, thereby realizing intelligent recognition of the posture of the target human body.

[0139] By performing the above operations, this solution addresses the technical problem that existing human posture recognition models, which rely solely on static or single-modal features, cannot effectively capture the spatial linkage and temporal evolution of human joints in complex scenes, resulting in low accuracy of human posture recognition results. It innovatively designs a joint angular velocity dynamic attention layer and hierarchical spatiotemporal graph convolution. By dynamically adjusting the feature channel weights based on human joint angular velocity and capturing the spatial linkage and temporal changes of target human movements layer by layer, it strengthens the ability to model temporal movements, significantly improves the modeling accuracy of temporal features and the adaptability to dynamic scenes, effectively enhances the accuracy of human posture recognition, and achieves accurate recognition of target human postures in high-speed motion and complex action scenes.

[0140] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0141] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0142] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. An artificial intelligence-based human body posture intelligent recognition system, characterized in that: The system comprises a data collection module, an image data optimization module, a human posture key point extraction module and a human posture intelligent recognition module. The data collection module is specifically configured to obtain human posture recognition original data by collecting data. The image data optimization module is specifically configured to obtain human posture optimized image data by image data cleaning, image deduplication, image denoising and image normalization. The human posture key point extraction module is specifically configured to construct a human posture key point extraction model by shallow feature initialization extraction, human joint multi-scale feature calibration, human whole body limb correlation feature capture, human joint fusion feature extraction and human posture key point output, and perform model training, and finally input target posture recognition data into the trained model to obtain a target human posture key point set. The human joint multi-scale feature calibration is specifically configured to input the human posture shallow feature map into 8 human joint multi-scale feature enhancement blocks in series to obtain a human joint multi-scale enhanced feature map. The human joint multi-scale feature enhancement block specifically comprises the following steps: Human joint multi-scale parallel feature extraction is specifically configured to extract human joint multi-scale parallel features by using G parallel convolution layers, and perform splicing operation on the G-layer output to obtain a multi-scale fusion feature map. Joint channel feature calibration is specifically configured to perform global average pooling and global maximum pooling operations on the multi-scale fusion feature map, and input the pooling results into a fully connected layer and a ReLU activation function respectively to obtain two channel feature vectors, perform splicing operation on the two channel feature vectors, and generate joint channel attention weights through a fully connected layer and a Sigmoid activation function; perform element-wise multiplication operation on the channel attention weights and the multi-scale fusion feature map to obtain a joint channel feature map. Joint pixel feature calibration is specifically configured to perform convolution operation on the joint channel feature map, and generate pixel attention weights through a Sigmoid activation function; perform element-wise multiplication operation on the pixel attention weights and the joint channel feature map to obtain a joint pixel calibration feature map. Residual connection feature fusion is specifically configured to splice the joint pixel calibration feature map and the human joint multi-scale enhanced feature map output by the previous human joint multi-scale feature enhancement block to obtain the human joint multi-scale enhanced feature map of the current human joint multi-scale feature enhancement block. The human posture intelligent recognition module is specifically configured to obtain target human posture intelligent recognition results by establishing a human posture recognition model, training the recognition model, inputting the target human posture key point set into the trained human posture recognition model; the establishment of the human posture recognition model comprises constructing a human motion space graph convolution layer, constructing a joint angular velocity dynamic attention layer, constructing a human posture time series convolution layer and constructing a human posture classification output layer. 2.The human body posture intelligent recognition system based on artificial intelligence according to claim 1, characterized in that: The human posture key point extraction module specifically comprises the following steps: Shallow feature initialization extraction is specifically configured to perform initial feature mapping on the image by convolution operation combined with a ReLU activation function to obtain a human posture shallow feature map. Human joint multi-scale feature calibration; The human body whole body limb correlation feature capturing specifically comprises: performing convolution operation on human joint multi-scale enhanced feature maps through four parallel dilated convolution layers; and performing splicing operation on the output features of the four dilated convolution layers to obtain a human body whole body correlation feature map. The human joint fusion feature extraction specifically comprises: performing convolution operation on the human body posture shallow layer feature map, and performing element-by-element multiplication operation on the human body whole body correlation feature map to obtain a human joint fusion feature map. The human body posture key point output is specifically that a convolution operation is performed on a human body joint fusion feature map, and then an output value is mapped to An interval is generated, a human body key point confidence map is generated, and for a channel corresponding to the k joint, all pixels in the channel are traversed, the pixel position with the maximum confidence is selected as the coordinate of the joint, the coordinates of N human body joints and the corresponding confidence are integrated, and a human body posture key point set is obtained, wherein k represents the index of the joint. The model is constructed and trained, specifically: a human body posture key point extraction model is constructed through shallow feature initialization extraction, human joint multi-scale feature calibration, human body whole body limb correlation feature capturing, human joint fusion feature extraction, and human body posture key point output; the model is trained based on reference posture recognition data of the human body posture optimization image data as training data, to obtain a trained human body posture key point extraction model. The target human body posture key point extraction specifically comprises: inputting target posture recognition data of the human body posture optimization image data into the trained human body posture key point extraction model to obtain a target human body posture key point set. 3.The human posture intelligent recognition system based on artificial intelligence according to claim 1, characterized in that: The human body posture intelligent recognition module specifically comprises the following steps: establishing a human body posture recognition model; training the model, specifically: training the model based on reference posture recognition data in the human body posture optimization image data as training data, to obtain a trained human body posture recognition model; intelligently recognizing a target human body posture, specifically: inputting the target human body posture key point set into the trained human body posture recognition model to obtain a target human body posture intelligent recognition result, to realize intelligent recognition of the target human body posture. 4.The human posture intelligent recognition system based on artificial intelligence according to claim 1, characterized in that: The human body posture recognition model is established, specifically comprising the following steps: constructing a human body motion space graph convolution layer; constructing a joint angular velocity dynamic attention layer; constructing a human body posture time series convolution layer, specifically: splicing enhanced spatial correlation features into a time series feature sequence according to a time dimension, and performing time series convolution operation by using one-dimensional convolution, and combining a ReLU activation function to strengthen nonlinear expression, to obtain human body posture time series features; constructing a human body posture classification output layer, specifically: performing global average pooling operation on the human body posture time series features, and performing feature dimension conversion through two fully connected layers, and finally outputting a class probability through a Softmax activation function, and selecting the class with the largest probability as a human body posture output result. 5.The human pose intelligent recognition system based on artificial intelligence according to claim 1, characterized in that: The human body motion space graph convolution layer is constructed, specifically comprising the following steps: human body posture space-time graph establishment, specifically: obtaining a graph node set, a spatial relationship edge set, and a normalized spatial adjacency matrix set by constructing a graph node, constructing a spatial relationship edge, and constructing a spatial adjacency matrix; the spatial adjacency matrix is constructed, specifically: constructing an initial adjacency matrix based on the spatial relationship edge set, and dividing the initial adjacency matrix into a centripetal joint sub-matrix, a centrifugal joint sub-matrix, and a joint self-connection sub-matrix according to the straight line distance from the joint to the human body center of gravity, and then performing normalization processing on the sub-matrices to obtain a normalized spatial adjacency matrix set; The spatial graph convolution feature extraction comprises the following steps: 6.The human pose intelligent recognition system based on artificial intelligence according to claim 1, characterized in that: The joint angle velocity dynamic attention layer is constructed; The method comprises the following steps: The joint angle velocity time average is calculated, specifically by using the joint coordinates of consecutive frames, the angular velocity of the kth joint at the tth frame is calculated, and the joint angular velocity average of the kth joint in the T frame action period is calculated, wherein T represents the total number of preset action period frames, and t represents the frame index; The channel response intensity is calculated, specifically by performing a global average pooling operation on the spatial graph convolution feature according to the joint and time sequence dimensions to obtain the channel response intensity; The dynamic weight generation comprises the following steps: The feature calibration comprises the following steps: 7.The human pose intelligent recognition system based on artificial intelligence according to claim 1, characterized in that: The data collection module, specifically through data collection operations, obtains human posture recognition original data; the human posture recognition original data includes reference posture recognition data and target posture recognition data; the reference posture recognition data and the target posture recognition data both include human posture image data, and the reference posture recognition data further includes reference human joint key point annotation data and reference posture annotation data. 8.The human posture intelligent recognition system based on artificial intelligence according to claim 1, characterized in that: The image data optimization module comprises the following steps: The image data cleaning, specifically by using an image quality detection algorithm, removes incomplete or low-quality images in the human posture image data; The image deduplication, specifically by using a similarity detection algorithm, compares and filters consecutive frames or similar image frames, and deletes duplicate images; The image denoising, specifically by using a Gaussian filtering technique to smooth the image; The image normalization, specifically by using a standardization process to adjust the pixel value range of the image to a preset standard interval. The image normalization, specifically by using a standardization process to adjust the pixel value range of the image to a preset standard interval.

Images