Human body posture anomaly detection method based on multi-branch space-time modeling

Abstract

A human body posture anomaly detection method based on multi-branch space-time modeling comprises the following steps: firstly, designing a multi-branch local-global feature fusion module, and carrying out refined posture feature modeling on parts such as four limbs and trunk of a human body by constructing local feature extraction branches for each sub-region of the body; dynamic fusion of local features and global features is realized in combination with an adaptive weighting mechanism, and a multi-level feature representation system is formed, so that the detection precision of sudden local anomalies and global attitude anomalies is synchronously improved; meanwhile, a time sequence attention enhancement module is provided, a time sequence attention mechanism fusing a bidirectional circulation unit is adopted, forward and backward time sequence context information is parallelly integrated, the modeling capacity of the model for long-time-span attitude dynamic change is enhanced, and the recognition robustness of continuous abnormal attitudes is remarkably improved. The method shows excellent performance in local sudden anomaly and persistent anomaly detection tasks.

Description

Technical Field

[0001] The present invention relates to a method for detecting anomalies in human posture, in particular to a method for detecting anomalies in human posture based on multi-branch spatiotemporal modeling, and belongs to the technical field of video anomaly detection. Background Art

[0002] With the rapid development of intelligent technologies, abnormal behavior detection has become a key technical means for improving the effectiveness of public safety management. Traditional monitoring methods face significant challenges in complex scenarios, such as crowded public places (including tourist attractions, transportation hubs, and commercial centers) and complex industrial environments (such as production workshops, warehouses, and energy facilities). Issues such as dynamic crowd flow, occlusion, and behavioral diversity in public settings, as well as mechanical noise, complex backgrounds, and equipment interference in industrial environments, all pose significant challenges to anomaly detection. Traditional methods based on manual monitoring and simple motion detection algorithms not only require significant human resources but also exhibit significant limitations when faced with complex tasks such as multi-target tracking and identifying sudden abnormal events. Particularly in open, highly dynamic environments, traditional methods often have detection blind spots, making it difficult to achieve all-weather, all-round security monitoring. This can lead to missed detection of critical abnormal behaviors, posing a safety hazard.

[0003] Existing anomaly detection methods are mainly divided into two categories: video frame-based visual feature detection and behavioral feature-based detection. Video frame-based visual feature detection methods typically identify abnormal behavior by detecting visual information such as inter-frame differences and optical flow. Although these methods perform well in certain application scenarios, their performance is unstable in scenarios with a large number of abnormal events. Behavioral feature-based detection methods detect abnormal behavior by analyzing people's behavioral trajectories and posture changes. These methods are highly robust to environmental changes, but using these behavioral feature extraction methods in public scenarios often leads to complex image processing algorithms, high computational complexity, and poor real-time performance.

[0004] With the development of deep learning technology, behavioral feature extraction methods based on skeleton points have gradually attracted attention. Skeleton points are key points on the human body that can accurately reflect a person's posture and gestures. Skeleton points can effectively extract behavioral features. Skeleton-based video anomaly detection, in public scenes, effectively detects abnormal behavior by modeling normal behavior, compared to traditional anomaly detection methods that require large amounts of abnormal data. However, existing methods are ineffective for subtle body gestures and continuous abnormal gestures, and are easily misclassified as normal behavior. Summary of the Invention

[0005] The purpose of the present invention is to provide a human posture anomaly detection method based on multi-branch spatiotemporal modeling, which, through local feature enhancement and temporal modeling optimization, improves the accuracy of sudden limb anomaly detection while effectively improving the recognition ability of persistent abnormal posture.

[0006] In order to achieve the above object, the present invention provides a method for detecting human posture anomaly based on multi-branch spatiotemporal modeling, comprising the following steps:

[0007] S1: Use the camera to collect video data, and use the skeleton point extraction network to extract the skeleton point graph data X = {B, T, V, C} from the video sequence of the video data, where B represents the number of batch samples, T represents the number of video frames, V represents the number of key points, C represents the key point coordinates (x, y) and the adjacency matrix A representing the connection relationship of the skeleton points. The information representing the coordinates of the skeleton points in the extracted skeleton point graph data is expanded by one dimension to obtain the skeleton point coordinate information. For the skeleton point graph data, there are 17 key points V 17 ={P0, P1, ...P 16}, newly added neck key point P 17 It is the midpoint of the left shoulder P5 and the right shoulder P6. Then the skeleton point map data is activated normalized and channel permuted to obtain the skeleton point data X′;

[0008] S2: Spatially segment the normalized and channel-permuted skeleton point data X′ to obtain key point data representing the upper and lower body. A spatial graph convolutional network is constructed for each set of key points, capturing the local spatial features of the upper and lower body through independent graph structure modeling. Subsequently, the local spatial features of the upper and lower body are weightedly fused with the global spatial features extracted from the skeleton point data X′ to form a joint feature representation with multi-level spatial expression capabilities.

[0009] S3: Adopts a bidirectional temporal modeling strategy. For the bidirectional recurrent unit, the temporal dependencies of the skeleton point data X′ are calculated through forward and backward recurrent neural networks, respectively, to obtain a bidirectional temporal score that contains long-range temporal context information. For the attention module, local one-dimensional convolution and global temporal pooling operations are used in parallel in the temporal dimension to establish short-term receptive fields and global receptive fields, respectively, extracting local abnormal patterns and global salient features to capture instantaneous abnormal posture patterns and overall motion trends. Finally, the bidirectional temporal score, local abnormal patterns, and global salient features are fused to generate a comprehensive attention score.

[0010] S4: The attention score calculated in step S3 is weightedly fused with the joint feature representation extracted in step S2 to obtain the features used to calculate the anomaly score, thereby enhancing the expressive power of key spatiotemporal features. Subsequently, the features after weighted fusion of steps S3 and S2 are mapped to the latent space through the normalized flow architecture to obtain an updated pose sequence, and the anomaly score is calculated based on the probability density estimation. If the score exceeds the preset detection threshold, the current input skeleton point map data X is determined to be abnormal behavior.

[0011] The specific steps of the fusion feature generation method in step S2 of the present invention are as follows:

[0012] S21: For the skeleton point data X′ processed in step S1, in order to distinguish local posture features, the skeleton point data X′ is divided into two sub-parts, the upper body and the lower body, and features are extracted from each sub-part;

[0013] X′1,X′2=split(X′,dim=-2)

[0014] Among them: X′1 represents the joint feature of the upper body;

[0015] X′2 represents the joint features of the lower body;

[0016] S22: In the process of extracting local posture features of the upper and lower body, graph convolution operations (STGCN) are performed on the postures of the upper and lower body respectively to extract the spatiotemporal features of the local postures. For the upper body part, its corresponding bone connection adjacency matrix A1 is defined, and the upper body features are extracted through graph convolution operations (STGCN):

[0017] H′1=STGCN1(X′1,A1·I1)

[0018] Similarly, for the lower body, define its corresponding bone connection adjacency matrix A2, and extract the lower body features through graph convolution operation (STGCN):

[0019] H′2=STGCN1(X′2,A2·I2)

[0020] Where: I1 and I2 represent the corresponding important matrices, respectively, controlling the important weights of local bone connections;

[0021] S23: After extracting local features through the graph convolution operation (STGCN), the upper body feature H′1 and the lower body feature H′2 are concatenated to form the overall local posture feature representation:

[0022] H concat =concat(H′1,H′2,dim=-2)

[0023] Then the multi-layer perceptron is used to further adjust the feature dimension to generate the fused local feature representation H fused :

[0024] H fused =MLP(H concat );

[0025] S24: Global pose features are extracted through GCN. In order to combine local pose features with global pose features, weighted fusion is used. By learning weight parameters, the graph convolutional network GCN can be adjusted according to the importance of local pose features and global pose features in different scenarios:

[0026] H′=α·GCN(X)+β·H fused

[0027] Among them: α and β are learnable weight parameters, α + β = 1, ensuring that the fused features remain intact;

[0028] H′ represents the overall feature after the fusion of local posture features and global posture features.

[0029] This approach can dynamically adjust the contribution of local and global features, making the model more flexible for different scenarios.

[0030] The specific steps of the temporal attention generation method in step S3 of the present invention are as follows:

[0031] S31: For the input skeleton point data X′∈RB ×C×T×V , through a fully connected layer to transform the dimension of the skeleton point data X′ and adjust it to the intermediate representation X L ∈R B×C×T×M , where M represents the number of neurons in the bidirectional recurrent unit, where M is the number of hidden units set in the bidirectional recurrent neural network. Then, for each independent time step t, the corresponding feature slice x(t)∈R is extracted B×C×1×M , and input them into the forward propagation GRU unit and the backward propagation GRU unit for temporal feature modeling. These two recurrent neural networks in opposite directions capture the temporal context dependency of the sequence data. The forward GRU processes the historical information from the beginning to the current moment, while the backward GRU processes the future information from the end to the current moment, obtaining the hidden information in two directions. Finally, by fusing the hidden information in these two directions, a new memory information F is formed. L ,Right now:

[0032] F L =[l(1),l(2),…,l(T)] T ∈R B×C×T×M

[0033] in: Represents the memory information of the t-th frame;

[0034] and They are the forward and backward cycle unit outputs respectively;

[0035] S32: For the attention calculation module, the skeleton point image data X is first upgraded, and then a one-dimensional convolution is used to establish a local receptive field in the time dimension to capture short-term abnormal patterns; at the same time, the maximum pooling is used to extract the global significant features in the time series. The short-term abnormal features and the global significant features are fused after the dimensionality reduction full connection layer transformation, and finally the attention weight F is generated through the Sigmoid function. A , the specific formula is expressed as:

[0036] F A =σ(W2Conv1d(W1X)+W3MaxPool(W1X))∈R B×C×T×V

[0037] Among them: W1 is the dimension-raising fully connected layer;

[0038] W2 and W3 represent dimension reduction fully connected layers;

[0039] σ refers to the sigmoid function;

[0040] The memory information and attention score are fused by element-by-element multiplication, and the dimension is adjusted to obtain the comprehensive attention score through residual connection. The comprehensive attention score is then weighted to the skeleton point data X′ to obtain the skeleton attention feature X''.

[0041] The specific steps of the anomaly score calculation method in step S4 of the present invention are as follows:

[0042] S41: Fuse the skeleton attention feature X'' from step S3 with the overall feature H' from S2 to obtain the feature tensor G', and divide the feature tensor G' into two subsets: shift and scale:

[0043] G′→(shift, scale);

[0044] S42: First, calculate the global average pooling value of the feature tensor G′, and add 1 after processing it with the tanh activation function to obtain the dynamic temperature coefficient. On this basis, divide the original scaling parameter scale by the temperature coefficient τ and send it to the Sigmoid function for nonlinear transformation, then multiply it by 2 and add a minimum value 1e -6 The bias term composed of the learnable noise suppression parameter β finally obtains the optimized scaling factor:

[0045] τ=1+tanh(GlobalAvgBool(G′))

[0046]

[0047] S43: Using residual feature calibration:

[0048] G”=(G′+α·shift)⊙scale_adj

[0049] α is the adaptive attenuation factor: α = exp(-‖shift‖2)

[0050] S43: log-determinant′ calculation, update the determinant logdet by taking the logarithm of scale and summing over the specified dimension:

[0051]

[0052] S44: Use G ” The probability density score representing the anomaly is calculated using logdet′, and the threshold is calculated using a dynamic threshold mechanism;

[0053] threshold=Percentile(anomaly_score_history,95%)+δ·MAD

[0054] score=1 / 2(G”^TΣ^(-1)G”+log|Σ|)-logdet′

[0055] MAD represents the mean absolute deviation of the score in the current window;

[0056] δ represents the sensitivity coefficient;

[0057] The finally calculated anomaly score is compared with the threshold. When the anomaly score is greater than the threshold, it indicates that an anomaly exists in the current skeleton point image data X.

[0058] Compared with the prior art, the skeleton point-based detection method of the present invention focuses on the local posture of a person and continuous abnormal behavior. It uses the key point data of the human body for analysis and is independent of the scene background. Therefore, it has stronger robustness to environmental changes. In open and dynamically changing environments such as public places, skeleton point detection can more stably identify human behavior. At the same time, the use of the skeleton point method can capture the specific posture of the human body in more detail, especially in complex crowd scenes, and can effectively distinguish different types of behavior. For example, abnormal behaviors such as fighting, running, and falling can be identified in their specific form through skeleton point data, without being easily interfered with by the background. Moreover, for dynamic scenes such as public places, skeleton point-based anomaly detection only needs to analyze the position and relative changes of the key points of the human body, which significantly reduces the amount of calculation compared to traditional image pixel-level processing. Therefore, in scenes with dense crowds and complex environments, the skeleton point method can provide higher real-time performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 A flowchart of the invention;

[0060] Figure 2 Strengthen the global posture flow chart for local posture;

[0061] Figure 3 This is the temporal attention flowchart. DETAILED DESCRIPTION

[0062] In high-risk environments like industrial manufacturing, construction, and logistics, workers can experience safety accidents due to sudden, illegal operations (such as accidentally touching machinery or engaging in unsafe postures) or persistent fatigue (such as prolonged bending or repetitive movements). Traditional monitoring relies on manual inspections or simple rules, making it difficult to accurately identify complex posture risks. The proposed human posture anomaly detection method based on multi-branch spatiotemporal modeling can effectively address both of these issues.

[0063] The present invention will be further described below with reference to the accompanying drawings.

[0064] like Figure 1-Figure 3 As shown, a method for detecting abnormal human posture based on multi-branch spatiotemporal modeling includes the following steps:

[0065] S1: Video data is collected by cameras installed in industrial production workshops, and the target detection algorithm is used to extract the personnel information in the video. The skeleton point extraction network is used to extract the skeleton point graph data X = {B, T, V, C} from the video sequence of the video data, where B represents the number of batch samples, T represents the number of video frames, V represents the number of key points, C represents the key point coordinates (x, y) and the adjacency matrix A representing the connection relationship of the skeleton points. The information representing the skeleton point coordinates in the extracted skeleton point graph data is expanded by one dimension to obtain the skeleton point coordinate information. For the skeleton point graph data, there are 17 key points V 17 ={P0, P1, ...P 16}, newly added neck key point P 17 It is the midpoint between the left shoulder P5 and the right shoulder P6:

[0066] Expand to

[0067] For the skeleton point data obtained after the processed expansion dimension, activation normalization is performed, and a linear transformation is performed on the dimension representing the key point coordinate C to make its mean zero and standard deviation one. It contains two parameters: translation parameter and scaling parameter. These two parameters can be adjusted through learning:

[0068] X′=s·X+b

[0069] Where: s represents the learnable scaling parameter, b represents the learnable translation parameter, and X′ represents the normalized bone point data; in order to calculate the inverse operation:

[0070]

[0071] Another output of activation normalization, log determinant, is an important indicator for measuring the change in transformation volume and is used to calculate the log-likelihood of the change in data distribution:

[0072]

[0073] The normalized skeleton point data X′ is subjected to spatial channel exchange and temporal sequence permutation. After permuting T:{t1, t2, …tn} and C:{x, y} in the skeleton point graph data, we obtain:

[0074]

[0075] S2: Spatially segment the normalized and channel-permuted skeleton point data X′ to obtain key point data representing the upper and lower body, dividing it into two sets of key point sets: the upper and lower body. Spatial graph convolutional networks are constructed for each set of key point sets, capturing the local spatial features of the upper and lower body through independent graph structure modeling. Subsequently, the local spatial features of the upper and lower body are weightedly fused with the global spatial features extracted from the skeleton point data X′ to form a joint feature representation with multi-level spatial expression capabilities.

[0076] The specific steps of the fusion feature generation method in step S2 are as follows:

[0077] S21: For the skeleton point data X′ processed in step S1, in order to distinguish local posture features, the posture sequence is divided into two sub-parts, the upper body and the lower body, and features are extracted for each part;

[0078] X′1,X′2=split(X′,dim=-2)

[0079] Among them: X′1 represents the joint feature of the upper body;

[0080] X′2 represents the joint features of the lower body;

[0081] S22: In the process of extracting local posture features of the upper and lower body, graph convolution operations (STGCN) are performed on the postures of the upper and lower body respectively to extract the spatiotemporal features of the local postures. For the upper body part, its corresponding bone connection adjacency matrix A1 is defined, and the upper body features are extracted through graph convolution operations (STGCN):

[0082] H′1=STGCN1(X′1,A1·I1)

[0083] Similarly, for the lower body, define its corresponding bone connection adjacency matrix A2, and extract the lower body features through graph convolution operation (STGCN):

[0084] H′2=STGCN1(X′2,A2·I2)

[0085] Where: I1 and I2 represent the corresponding important matrices, respectively, controlling the important weights of local bone connections;

[0086] S23: After extracting local features through the graph convolution operation (STGCN), the upper body feature H′1 and the lower body feature H′2 are concatenated to form the overall local posture feature representation:

[0087] H concat =concat(H′1,H′2,dim=-2)

[0088] Then the multi-layer perceptron is used to further adjust the feature dimension to generate the fused local feature representation H fused :

[0089] H fused =MLP(H concat );

[0090] S24: Global pose features are extracted through GCN. In order to combine local pose features with global pose features, weighted fusion is used. By learning weight parameters, the graph convolutional network GCN can be adjusted according to the importance of local pose features and global pose features in different scenarios:

[0091] H′=α·GCN(X)+β·H fused

[0092] Among them: α and β are learnable weight parameters, α + β = 1, ensuring that the fused features remain intact;

[0093] This approach can dynamically adjust the contribution of local and global features, making the model more flexible for different scenarios.

[0094] S3: To address the persistent fatigue postures that workers may experience (such as prolonged bending and repetitive movements), a bidirectional temporal modeling strategy is adopted. For the bidirectional recurrent unit, the temporal dependencies of the skeleton point data X′ are calculated through forward and backward recurrent neural networks, respectively, to obtain a bidirectional temporal score that contains long-range temporal context information. For the attention module, local one-dimensional convolution and global temporal pooling operations are used in parallel in the temporal dimension to establish short-term receptive fields and global receptive fields, respectively. Local abnormal patterns and global salient features are extracted to capture instantaneous abnormal posture patterns and overall motion trends. Finally, the bidirectional temporal score, local abnormal patterns, and global salient features are fused to generate a comprehensive attention score.

[0095] The specific steps of the temporal attention generation method in step S3 are as follows:

[0096] S31: For the input skeleton point data X′∈R B×C×T×V , through a fully connected layer to transform the dimension of the skeleton point data X′ and adjust it to the intermediate representation X L ∈R B×C×T×M , where M represents the number of neurons in the bidirectional recurrent unit, where M is the number of hidden units set in the bidirectional recurrent neural network. Then, for each independent time step t, the corresponding feature slice x(t)∈R is extracted B×C×1×M, and input them into the forward propagation GRU unit and the backward propagation GRU unit for temporal feature modeling. These two recurrent neural networks in opposite directions capture the temporal context dependency of the sequence data. The forward GRU processes the historical information from the beginning to the current moment, while the backward GRU processes the future information from the end to the current moment, obtaining the hidden information in two directions. Finally, by fusing the hidden information in these two directions, a new memory information F is formed. L ,Right now:

[0097] F L =[l(1),l(2),…,l(T)] T ∈R B×C×T×M

[0098] in: Represents the memory information of the t-th frame;

[0099] and They are the forward and backward cycle unit outputs respectively;

[0100] S32: For the attention calculation module, the skeleton point image data X is first upgraded, and then a one-dimensional convolution is used to establish a local receptive field in the time dimension to capture short-term abnormal patterns; at the same time, the maximum pooling is used to extract the global significant features in the time series. The short-term abnormal features and the global significant features are fused after the dimensionality reduction full connection layer transformation, and finally the attention weight F is generated through the Sigmoid function. A , the specific formula is expressed as:

[0101] F A =σ(W2Conv1d(W1X)+W3MaxPool(W1X))∈R B×C×T×V

[0102] Among them: W1 is the dimension-raising fully connected layer;

[0103] W2 and W3 represent dimension reduction fully connected layers;

[0104] σ refers to the sigmoid function;

[0105] The memory information and attention score are fused by element-by-element multiplication, and the dimension is adjusted to obtain the comprehensive attention score through residual connection. The comprehensive attention score is then weighted to the skeleton point data X′ to obtain X′′.

[0106] S4: The attention score calculated in step S3 is weightedly fused with the joint feature representation extracted in step S2 to obtain the features used to calculate the anomaly score, thereby enhancing the expressive power of key spatiotemporal features. Subsequently, the features after weighted fusion of steps S3 and S2 are mapped to the latent space through the normalized flow architecture to obtain an updated pose sequence, and the anomaly score is calculated based on the probability density estimation. If the score exceeds the preset detection threshold, the current input skeleton point map data X is determined to be abnormal behavior.

[0107] The specific steps of the anomaly score calculation method in step S4 are as follows:

[0108] S41: Fuse the skeleton attention feature X'' from step S3 with the overall feature H' from S2 to obtain the feature tensor G', and divide the feature tensor G' into two subsets: shift and scale:

[0109] G′→(shift, scale);

[0110] S42: First, calculate the global average pooling value of the feature tensor G′, and add 1 after processing it with the tanh activation function to obtain the dynamic temperature coefficient. On this basis, divide the original scaling parameter scale by the temperature coefficient τ and send it to the Sigmoid function for nonlinear transformation, then multiply it by 2 and add a minimum value 1e -6 The bias term composed of the learnable noise suppression parameter β finally obtains the optimized scaling factor:

[0111] τ=1+tanh(GlobalAvgBool(G′))

[0112]

[0113] S43: Using residual feature calibration:

[0114] G″=(G′+α·shift)⊙scale_adj

[0115] α is the adaptive attenuation factor: α = exp(-‖shift‖2)

[0116] S43: log-determinant′ calculation, update the determinant logdet by taking the logarithm of scale and summing over the specified dimension:

[0117]

[0118] S44: Use G″ and logdet′ to calculate the probability density score representing the anomaly, and calculate the threshold through the dynamic threshold mechanism;

[0119] threshold=Percentile(anomaly_score_history,95%)+δ·MAD

[0120] score=1 / 2(G″^TΣ^(-1)G″+log|Σ|)-logdet′

[0121] MAD represents the mean absolute deviation of the score in the current window;

[0122] δ represents the sensitivity coefficient;

[0123] The finally calculated anomaly score is compared with the threshold. When the anomaly score is greater than the threshold, it indicates that an anomaly exists in the current skeleton point image data X.

Claims

1. A method for detecting abnormal human posture based on multi-branch spatiotemporal modeling, characterized in that: The following steps are involved: S1: Use the camera to collect video data, and use the skeleton point extraction network to extract the skeleton point graph data X = {B, T, V, C} from the video sequence of the video data, where B represents the number of batch samples, T represents the number of video frames, V represents the number of key points, C represents the key point coordinates (x, y) and the adjacency matrix A representing the connection relationship of the skeleton points. The information representing the coordinates of the skeleton points in the extracted skeleton point graph data is expanded by one dimension to obtain the skeleton point coordinate information. For the skeleton point graph data, there are 17 key points V 17 ={P0, P1, ...P 16 }, newly added neck key point P 17 It is the midpoint of the left shoulder P5 and the right shoulder P6. Then the skeleton point map data is activated normalized and channel permuted to obtain the skeleton point data X'; S2: Spatially segment the normalized and channel-permuted skeleton point data X' to obtain key point data representing the upper and lower body. A spatial graph convolutional network is constructed for each set of key points, capturing the local spatial features of the upper and lower body through independent graph structure modeling. Subsequently, the local spatial features of the upper and lower body are weightedly fused with the global spatial features extracted from the skeleton point data X' to form a joint feature representation with multi-level spatial expression capabilities. S3: Adopts a bidirectional temporal modeling strategy. For the bidirectional recurrent unit, the temporal dependencies of the skeleton point data X' are calculated through forward and backward recurrent neural networks, respectively, to obtain a bidirectional temporal score that contains long-range temporal context information. For the attention module, local one-dimensional convolution and global temporal pooling operations are used in parallel in the temporal dimension to establish short-term receptive fields and global receptive fields, respectively. Local abnormal patterns and global salient features are extracted to capture instantaneous abnormal posture patterns and overall motion trends. Finally, the bidirectional temporal score, local abnormal patterns, and global salient features are fused to generate a comprehensive attention score. S4: The attention score calculated in step S3 is weightedly fused with the joint feature representation extracted in step S2 to obtain the features used to calculate the anomaly score, thereby enhancing the expressive power of key spatiotemporal features. Subsequently, the features after weighted fusion of steps S3 and S2 are mapped to the latent space through the normalized flow architecture to obtain an updated pose sequence, and the anomaly score is calculated based on the probability density estimation. If the score exceeds the preset detection threshold, the current input skeleton point map data X is determined to be abnormal behavior.

2. The method for detecting abnormal human posture based on multi-branch spatiotemporal modeling according to claim 1, characterized in that: The specific steps of the fusion feature generation method in step S2 are as follows: S21: For the skeleton point data X' processed in step S1, in order to distinguish local posture features, the skeleton point data X' is divided into two sub-parts, the upper body and the lower body, and features are extracted from each sub-part; X′1,X′2=split(X′,dim=-2) Among them: X′1 represents the joint feature of the upper body; X′2 represents the joint features of the lower body; S22: In the process of extracting local posture features of the upper and lower body, graph convolution operations STGCN are performed on the postures of the upper and lower body respectively to extract the spatiotemporal features of the local postures. For the upper body part, its corresponding bone connection adjacency matrix A1 is defined, and the upper body features are extracted through graph convolution operations STGCN: H′1=STGCN1(X′1,A1·I1) Similarly, for the lower body, define its corresponding bone connection adjacency matrix A2, and extract the lower body features through graph convolution operation STGCN: H′2=STGCN1(X′2,A2·I2) Where: I1 and I2 represent the corresponding important matrices, respectively, controlling the important weights of local bone connections; S23: After the graph convolution operation STGCN extracts local features, the upper body feature H′1 and the lower body feature H′2 are concatenated to form the overall local posture feature representation: H concat =concat(H′1,H′2,dim=-2) Then the multi-layer perceptron further adjusts the feature dimension to generate the fused local feature representation H fused : H fused =MLP(H concat ); S24: Global pose features are extracted through GCN. In order to combine local pose features with global pose features, weighted fusion is used. By learning weight parameters, GCN is adjusted according to the importance of local pose features and global pose features in different scenes: H′=α·GCN(X)+β·H fused Among them: α and β are learnable weight parameters; H′ represents the overall feature after the fusion of local posture features and global posture features.

3. The method for detecting abnormal human posture based on multi-branch spatiotemporal modeling according to claim 2, characterized in that: The specific steps of the temporal attention generation method in step S3 are as follows: S31: For the input skeleton point data X'∈R B×C×T×V , through a fully connected layer to transform the dimension of the skeleton point data X' and adjust it to the intermediate representation X L ∈R B×C×T×M , where M represents the number of neurons in the bidirectional recurrent unit, where M is the number of hidden units set in the bidirectional recurrent neural network. Then, for each independent time step t, the corresponding feature slice x(t)∈R is extracted B×C×1×M , and input them into the forward propagation GRU unit and the backward propagation GRU unit for temporal feature modeling. These two recurrent neural networks in opposite directions capture the temporal context dependency of the sequence data. The forward GRU processes the historical information from the beginning to the current moment, while the backward GRU processes the future information from the end to the current moment, obtaining the hidden information in two directions. Finally, by fusing the hidden information in these two directions, a new memory information F is formed. L ,Right now: F L =[l(1),l(2),…,l(T)] T ∈R B×C×T×M in: Represents the memory information of the t-th frame; and They are the forward and backward cycle unit outputs respectively; S32: For the attention calculation module, the skeleton point image data X is first upgraded, and then a one-dimensional convolution is used to establish a local receptive field in the time dimension to capture short-term abnormal patterns; at the same time, the maximum pooling is used to extract the global significant features in the time series. The short-term abnormal features and the global significant features are fused after the dimensionality reduction full connection layer transformation, and finally the attention weight F is generated through the Sigmoid function. A , the specific formula is expressed as: F A =σ(W2Conv1d(W1X)+W3MaxPool(W1X))∈R B×C×T×V Among them: W1 is the dimension-raising fully connected layer; W2 and W3 represent dimension reduction fully connected layers; σ refers to the sigmoid function; The memory information and attention score are fused by element-by-element multiplication, and the dimension is adjusted to obtain the comprehensive attention score through residual connection. The comprehensive attention score is then weighted to the skeleton point data X' to obtain the skeleton attention feature X''.

4. The method for detecting abnormal human posture based on multi-branch spatiotemporal modeling according to claim 1, characterized in that: The specific steps of the anomaly score calculation method in step S4 are as follows: S41: Fuse the skeleton attention feature X'' from step S3 with the overall feature H' from S2 to obtain the feature tensor G', and divide the feature tensor G' into two subsets: shift and scale: G′→(shift, scale); S42: First, the global average pooling value of the feature tensor G′ is calculated, and after being processed by the tanh activation function, 1 is added to obtain the dynamic temperature coefficient. On this basis, the original scaling parameter scale is divided by the temperature coefficient τ and then fed into the Sigmoid function for nonlinear transformation, and then multiplied by 2 and added with a minimum value 1e -6 The bias term composed of the learnable noise suppression parameter β finally obtains the optimized scaling factor: τ=1+tanh(GlobalAvgBool(G′)) S43: Using residual feature calibration: G”=(G’+α·shift)⊙scale_adj α is the adaptive attenuation factor: α = exp(-‖shift‖2) S43: log-determinant′ calculation, update the determinant logdet by taking the logarithm of scale and summing over the specified dimension: S44: Use G' and logdet' to calculate the probability density score representing the anomaly, and calculate the threshold through the dynamic threshold mechanism; threshold=Percentile(anomaly_score_history,95%)+δ·MAD score=1 / 2(G”^TΣ^(-1)G”+log|Σ|)-logdet' MAD represents the mean absolute deviation of the score in the current window; δ represents the sensitivity coefficient; The final calculated anomaly score is compared with the threshold. When the anomaly score is greater than the threshold, it indicates that there is an anomaly in the current sequence.

Images