A human action recognition method
Through the GRU timing processing module and 3D convolutional neural network combined with multi-level feature fusion and attention mechanism, a lightweight video human body motion recognition model is built, solving the problem of high timing information ignorance and computational complexity in the existing technology, and achieving high precision and efficient motion recognition.
Patent Information
- Application Number
- CN202211677196.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-12-26
- Publication Date
- 2025-08-22
- Estimated Expiration
- 2042-12-26
AI Technical Summary
The existing video human body movement recognition methods ignore timing information, resulting in low recognition accuracy and large network structure, making it difficult to operate efficiently on devices with limited resources.
The GRU timing processing module is used to combine 3D convolutional neural network, combined with multi-level feature fusion and spatial and channel attention mechanisms, and through sparse sampling and sparse feature extraction, a lightweight action recognition model is built to enhance timing features and important position information, and reduce the amount of parameters.
It improves the accuracy of video human movement recognition, reduces the computational complexity and resource requirements of the model, and improves the operation efficiency on limited resource equipment.
Smart Images
Figure CN116110124B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of data processing, and in particular to a method for recognizing human motion. Background Art
[0002] Video information is an important part of people's daily lives, and its processing and analysis methods have been widely studied. Video human action recognition has a wide range of applications, not only in action analysis, but also in advanced human-computer interaction, video surveillance, virtual reality, and other fields.
[0003] Traditional methods for human action recognition in videos are primarily based on manual feature extraction from local and global information. Improved Dense Trajectories (iDT), for example, is widely considered the most effective of these manual feature extraction methods. Existing deep learning-based methods for human action recognition include those combining long short-term memory networks (LSTMs) with convolutional neural networks (CNNs), methods based on graph convolution, and two-stream networks. However, each has its own shortcomings. LSTMs can only capture high-level actions at the top layer of the convolution layer and require long training times. Graph convolution methods can comprehensively study non-local joints, which are important for action recognition. In two-stream networks, optical flow only represents motion information between adjacent frames, making it difficult to model actions spanning longer timeframes.
[0004] 3D convolution-based human action recognition methods were developed from the study of 2D convolution on images. The main idea is to treat videos as images with an additional time dimension compared to images. This recognition method aims to address the problem of ignoring inter-frame motion information when processing sampled video frames. However, existing research models all ignore the impact of temporal information on the model, resulting in low recognition accuracy. Therefore, further improvement of existing technologies is needed. Summary of the Invention
[0005] The technical problem to be solved by the present invention is to provide a human motion recognition method with improved recognition accuracy and lightweight network structure in response to the above-mentioned existing technologies.
[0006] The technical solution adopted by the present invention to solve the above technical problems is: a human motion recognition method, characterized by comprising the following steps:
[0007] Step 1: Acquire multiple videos containing human body movements, and perform sparse sampling t frames on each video to obtain multiple sparsely sampled video frame sequences;
[0008] Step 2: Multiple video frame sequences are combined into a data set, and the data set is constructed into a training set and a test set; wherein each training sample in the training set includes a video frame sequence and a human action label corresponding to the video frame sequence;
[0009] Step 3: Construct an action recognition model and use the training set to train the constructed action recognition model to obtain a trained action recognition model.
[0010] The specific structure of the action recognition model is:
[0011] A GRU timing processing module includes a 3D convolutional neural network and k bidirectional single-layer GRU units connected to the 3D convolutional neural network. The input end of the 3D convolutional neural network serves as the input end of the GRU timing processing module, and the output ends of the k bidirectional single-layer GRU units serve as the output end of the GRU timing processing module.
[0012] The first spatial attention module has its input connected to the output of the GRU temporal processing module;
[0013] The first residual module, whose input is connected to the output of the first spatial attention module;
[0014] The second spatial attention module has its input connected to the output of the first residual module;
[0015] The second residual module has its input connected to the output of the second spatial attention module;
[0016] …
[0017] The input of the Nth spatial attention module is connected to the output of the N-1th residual module; N is a positive integer;
[0018] The input of the Nth residual module is connected to the output of the Nth spatial attention module;
[0019] The multi-level feature fusion module includes N-1 feature fusion modules, namely the first feature fusion module, the second feature fusion module...the N-1th feature fusion module, the output end of the N-th residual module is connected to the input end of the first feature fusion module, the output end of the N-1th residual module is connected to the output end of the first feature fusion module and then connected to the input end of the second feature fusion module; the output end of the N-2th residual module is connected to the output end of the second feature fusion module and then connected to the input end of the third feature fusion module...the output end of the first residual module and the output end of the N-1th feature fusion module are connected and then serve as the output end of the multi-level feature fusion module;
[0020] Segmented channel attention module, whose input is connected to the multi-level feature fusion module;
[0021] A feature extraction layer, comprising a pooling layer connected to the output of the segmented channel attention module and a fully connected layer connected to the pooling layer, wherein the fully connected layer is the output of the action recognition model;
[0022] Each of the above residual modules includes multiple identical residual blocks, and each residual block includes multiple convolutional layers connected in sequence. Each feature fusion module includes upsampling and convolution operations to make the dimension of its output end the same as the feature dimension of the output of its corresponding residual module, thereby enabling fusion.
[0023] The specific processing steps of the segmented channel attention module are:
[0024] The channel c corresponding to the feature map input to the segmented channel attention module is divided into M segments, and the number of channels in each of the M segments is obtained. The channel attention enhancement operation is performed on the M segments of the first feature map to obtain M segments of the second feature map. The M segments of the second feature map are concatenated and output along the channel dimension to obtain the output feature map of the segmented channel attention module.
[0025] Step 4: Randomly select one of the samples in the test set and input it into the action recognition model trained in step 3 to obtain the human action recognition result.
[0026] Specifically, the 3D convolutional neural network in step 3 includes a convolution layer with 3 input channels, 64 output channels, a convolution kernel size of 7*7*7, a step size of (1, 2, 2), and a padding of (3, 3, 3).
[0027] To avoid overfitting of the 3D convolutional neural network, the above 3 further includes between the 3D convolutional neural network and the k bidirectional single-layer GRU units: a normalization layer and an activation function layer connected after the convolution layer in the 3D convolutional neural network.
[0028] To enhance the spatial region of interest, the specific processing process of each spatial attention module in step 3 is as follows:
[0029] Step 3-1: Perform dimension transformation on the image output by the GRU time series processing module to obtain the first image X. The dimension of X is (c, t, h, w), where c, t, h, and w represent the channel dimension, the number of frames, and the height and width of each frame of video, respectively.
[0030] Step 3-2: compress the time dimension of the first image X to obtain a second image X', where the dimension of X' is transformed to (c, h, w);
[0031] Step 3-3: After the second image X' passes through two convolutional layers conv1 and conv2 respectively, it is then activated by the sigmoid function to generate a corresponding weight mask for each position and output weighted. Finally, it is added to the first image X to obtain the final output F, which has the same dimension as X.
[0032] To enhance the channel region of interest, the specific processing process of the channel attention enhancement operation in step 3 is as follows:
[0033] Step 3-a: Pool the spatiotemporal features except the channels in the M segments of the first feature map to compress their spatiotemporal dimensions to 1, thereby obtaining the first sub-map;
[0034] Step 3-b: Perform a one-dimensional convolution operation on the channel dimension of the first sub-image to obtain the second sub-image;
[0035] Step 3-c: Generate a corresponding weight mask for each channel position in the second sub-map through the activation function sigmoid operation and weighted output, and finally add the weighted output to the first feature map to obtain the second feature map.
[0036] Compared with the existing technology, the advantages of the present invention are: the temporal features of the video frame sequence are enhanced through the GRU unit, and the important position information of the video frame sequence is enhanced through the spatial attention mechanism; in addition, the multi-level feature fusion module can obtain the combined information of features at different levels in the video; finally, the segmented channel attention module can not only effectively extract important channel information and weaken areas of no interest, but also reduce the number of parameters and improve the running speed of the model. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] Figure 1 is a block diagram of an action recognition model in an embodiment of the present invention;
[0038] Figure 2 for Figure 1 Specific operation diagram of the spatial attention module;
[0039] Figure 3 for Figure 1 Specific operation diagram of the segmented channel attention module. DETAILED DESCRIPTION
[0040] The present invention will be described in further detail below with reference to the accompanying drawings and embodiments.
[0041] like Figure 1 As shown, the human motion recognition method in this embodiment includes the following steps:
[0042] Step 1: Acquire multiple videos containing human body movements, and sparsely sample t0 frames of each video to obtain multiple sparsely sampled video frame sequences;
[0043] Step 2: Multiple video frame sequences are combined into a data set, and the data set is constructed into a training set and a test set; wherein each training sample in the training set includes a video frame sequence and a human action label corresponding to the video frame sequence;
[0044] Step 3: Construct an action recognition model and use the training set to train the constructed action recognition model to obtain a trained action recognition model.
[0045] The specific structure of the action recognition model is:
[0046] A GRU timing processing module includes a 3D convolutional neural network and k bidirectional single-layer GRU units connected to the 3D convolutional neural network. The input end of the 3D convolutional neural network serves as the input end of the GRU timing processing module, and the output ends of the k bidirectional single-layer GRU units serve as the output end of the GRU timing processing module.
[0047] In this embodiment, the 3D convolutional neural network includes a convolution layer with 3 input channels, 64 output channels, a convolution kernel size of 7*7*7, a stride of (1, 2, 2), and a padding of (3, 3, 3). In addition, between the 3D convolutional neural network and the k bidirectional single-layer GRU units, a normalization layer and an activation function layer connected after the convolution layer in the 3D convolutional neural network are also included.
[0048] The first spatial attention module has its input connected to the output of the GRU temporal processing module;
[0049] The first residual module, whose input is connected to the output of the first spatial attention module;
[0050] The second spatial attention module has its input connected to the output of the first residual module;
[0051] The second residual module has its input connected to the output of the second spatial attention module;
[0052] …
[0053] The input of the Nth spatial attention module is connected to the output of the N-1th residual module; N is a positive integer;
[0054] The input of the Nth residual module is connected to the output of the Nth spatial attention module;
[0055] The multi-level feature fusion module includes N-1 feature fusion modules, namely the first feature fusion module, the second feature fusion module...the N-1th feature fusion module, the output end of the N-th residual module is connected to the input end of the first feature fusion module, the output end of the N-1th residual module is connected to the output end of the first feature fusion module and then connected to the input end of the second feature fusion module; the output end of the N-2th residual module is connected to the output end of the second feature fusion module and then connected to the input end of the third feature fusion module...the output end of the first residual module and the output end of the N-1th feature fusion module are connected and then serve as the output end of the multi-level feature fusion module;
[0056] Segmented channel attention module, whose input is connected to the multi-level feature fusion module;
[0057] A feature extraction layer, comprising a pooling layer connected to the output of the segmented channel attention module and a fully connected layer connected to the pooling layer, wherein the fully connected layer is the output of the action recognition model;
[0058] Each of the above residual modules includes multiple identical residual blocks, and each residual block includes multiple convolutional layers connected in sequence. Each feature fusion module includes upsampling and convolution operations to make the dimension of its output end the same as the feature dimension of the output of its corresponding residual module, thereby enabling fusion.
[0059] The specific processing steps of the segmented channel attention module are:
[0060] The channel c corresponding to the feature map input to the segmented channel attention module is divided into M segments to obtain a first feature map with a channel number of c / M in each segment; the channel attention enhancement operation is performed on the M segments of the first feature map to obtain M segments of the second feature map; the M segments of the second feature map are concatenated and output along the channel dimension to obtain the output feature map of the segmented channel attention module; in this embodiment, M=4, and the purpose of this segmentation is to reduce the amount of computation;
[0061] Step 4: Randomly select one of the samples in the test set and input it into the action recognition model trained in step 3 to obtain the human action recognition result.
[0062] like Figure 2 As shown, in this embodiment, the specific processing process of each spatial attention module is as follows:
[0063] Step 3-1: Perform dimension transformation on the image output by the GRU time series processing module to obtain the first image X. The dimension of X is (c, t, h, w), where c, t, h, and w represent the channel dimension, the number of frames, and the height and width of each frame of video, respectively.
[0064] Step 3-2: compress the time dimension of the first image X to obtain a second image X', where the dimension of X' is transformed to (c, h, w);
[0065] Step 3-3: After the second image X' passes through two convolutional layers conv1 and conv2 respectively, it is then activated by the sigmoid function to generate a corresponding weight mask for each position and output weighted. Finally, it is added to the first image X to obtain the final output F, which has the same dimension as X.
[0066] In this embodiment, the convolution layer conv1 is a convolution layer with 64 output channels, a convolution kernel size of 1*1*1, and a stride of 1. The convolution layer conv2 is a convolution layer with 64 input channels, a convolution kernel size of 1*1*1, and a stride of 1.
[0067] In addition, the channel attention enhancement operation mentioned above is roughly the same as the operation of the spatial attention module, such as Figure 3 As shown, the specific processing process of the channel attention enhancement operation in this embodiment is:
[0068] Step 3-a: Pool the spatiotemporal features except the channels in the M segments of the first feature map to compress their spatiotemporal dimensions to 1, thereby obtaining the first sub-map;
[0069] Step 3-b: Perform a one-dimensional convolution operation on the channel dimension of the first sub-image to obtain the second sub-image;
[0070] Step 3-c: Generate a corresponding weight mask for each channel position in the second sub-map through the activation function sigmoid operation and weighted output, and finally add the weighted output to the first feature map to obtain the second feature map.
[0071] In this embodiment, N=4, that is, there are 4 residual modules, Figure 1 The first residual module L1, the second residual module L2, the third residual module L3 and the fourth residual module L4 are shown in FIG.
[0072] L1 includes 3 identical residual blocks, each of which includes a convolution layer with 64 input channels, 64 output channels, a convolution kernel size of 1*1*1, and a stride of 1; a convolution layer with 64 input channels, 256 output channels, a convolution kernel size of 3*3*3, and a stride of 1; and a convolution layer with 256 input channels, 256 output channels, a convolution kernel size of 1*1*1, and a stride of 1.
[0073] L2 includes 4 identical residual blocks, each residual block includes a convolution layer with an input channel of 256, an output channel of 128, a convolution kernel size of 1*1*1, and a stride of 2; a convolution layer with an input channel of 128, an output channel of 128, a convolution kernel size of 3*3*3, a stride of 2, and padding of half the convolution kernel size; and a convolution layer with an input channel of 128, an output channel of 512, a convolution kernel size of 1*1*1, and a stride of 1.
[0074] L3 includes 23 identical residual blocks, each residual block includes a convolution layer with an input channel of 512, an output channel of 256, a convolution kernel size of 1*1*1, and a stride of 1; a convolution layer with an input channel of 256, an output channel of 256, a convolution kernel size of 3*3*3, a stride of 2, and padding of half the convolution kernel size; and a convolution layer with an input channel of 256, an output channel of 1024, a convolution kernel size of 1*1*1, and a stride of 1.
[0075] L4 includes 3 identical residual blocks, each residual block includes a convolution layer with an input channel of 1024, an output channel of 512, a convolution kernel size of 1*1*1, and a stride of 1; a convolution layer with an input channel of 512, an output channel of 512, a convolution kernel size of 3*3*3, a stride of 2, and padding of half the convolution kernel size; and a convolution layer with an input channel of 512, an output channel of 2048, a convolution kernel size of 1*1*1, and a stride of 1.
[0076] In this embodiment, the convolutional layer and nonlinear operations (normalized BN layer and activation function ReLU layer) in the 3D convolutional neural network are used to perform preliminary feature extraction on the input data information. The timing processing module based on the GRU unit is used in the initial stage of the module. After the video frame data for preliminary feature extraction is transformed in dimension, k bidirectional single-layer GRU units are used to obtain contextual semantic information at each moment and capture relationships with larger temporal distances in the video sequence.
[0077] In addition, the feature map after the GRU timing processing module is used to extract high-level information and enhance the spatial region of interest. Specifically, the information processed by the GRU timing processing module is transformed, and then the time dimension of the obtained features is compressed, and the key spatial information is retained by transforming the dimension; after two convolutions of conv1 and conv2 to reduce and increase the channels respectively, more high-level information is extracted; a corresponding weight mask is generated for each channel position and weighted output to enhance the spatial region of interest.
[0078] The spatial attention module is embedded before the residual layer. Each residual layer is composed of multiple residual blocks. In this embodiment, four residual layers are used, and the number of residual blocks contained in each residual layer is (3, 4, 23, 3) respectively. The output dimensions of L1, L2, L3, and L4 are (256, 8, 28, 28), (512, 4, 14, 14), (1024, 2, 7, 7), and (2048, 1, 4, 4), respectively. The output features of L4 are upsampled and convolved to make their dimensions consistent with the output dimensions of L3, and then the two are fused. The features obtained by feature fusion are upsampled and convolved to make their dimensions consistent with the output dimensions of L2, and the two are fused. The features after the second fusion are upsampled and convolved to make their dimensions consistent with the output dimensions of L1, and then the two are fused. The fused features are the final output features of the module.
[0079] Finally, based on the channel attention mechanism, the channels are segmented, and the channel attention enhancement operation is performed on each channel segment, and finally a joint output is performed based on the channel. The specific steps are as follows: the feature with c channels is divided into 4 segments, so that the number of channels in each segment is c / 4; the channel attention enhancement operation is performed on each feature segment. The specific method is to pool the spatiotemporal features except the channel so that its dimension becomes 1. The pooled features contain global information in the spatiotemporal dimension; after the dimensionality transformation, a one-dimensional convolution operation is performed on the channel dimension to retain the key channel information and extract more high-level information; the sigmoid function is used to generate a corresponding weight mask for each channel position and weighted output, and the channel area of interest in the c / 4 segment is enhanced. The output features of the four segments are connected and output along the channel dimension. The final classification prediction output is obtained by pooling and fully connecting the selected features.
[0080] The network is trained with the following parameters: epoch 15, batch size 16, learning rate 1e-3, L2 norm weight 1e-2, and input video sample size 16*3*112*112. The cross entropy loss function L is used for evaluation. cross Get the difference between the probability distribution obtained by the current training and the true distribution.
[0081]
[0082] Where N is the number of training samples, M is the number of classification categories, and y ic represents the label of sample i, p ic represents the predicted probability that sample i belongs to class c.
[0083] This embodiment uses the UCF101 dataset and the HMDB51 dataset as training and test datasets. The UCF101 dataset contains a total of 13,320 videos and 101 categories. It mainly includes human-object interaction, body movements, human-to-human interaction, playing musical instruments, and sports. The HMDB51 dataset contains a total of 6,766 videos and 51 categories. It mainly includes general facial movements, facial movements manipulated by objects, general body movements, body movements interacting with objects, and body movements interacting with people. In the UCF101 dataset, the recognition accuracy rate obtained by the experimental test according to the method of this embodiment is 96.42%. In the HMDB51 dataset, the recognition accuracy rate obtained by the experimental test according to the method of this embodiment is 71.08%.
[0084] Compared with other existing algorithms on the UCF101 dataset. Compared with the traditional manual human action recognition method iDT, the present invention has improved by 10.02%. Compared with the Two-stream network, the model proposed in this invention has improved by about 8.4% on the test set. Compared with the C3D network, the algorithm proposed in this invention has improved by about 11% on this basis. Compared with the 3DRseNet50-CS network, the model proposed in this article has improved by about 6.5% on the test set. Compared with the T3D algorithm, the model proposed in this article has improved by 3.2% on the test set. Compared with ResNeXt-101, the model proposed in this article has improved by 1.92% on the test set. Compared with the Improved DenseNet algorithm, the model proposed in this article has improved by 1.9% on the test set. Experimental results show that the improved network of this invention can achieve better recognition effect on the UCF101 dataset.
[0085] Comparison with other existing algorithms on the HMDB51 dataset. Compared with the iDT algorithm, the algorithm proposed in this paper improves by about 13% on this dataset. Compared with the Two-stream algorithm, the model proposed in this paper improves by about 10% on the test set. Compared with the 3DRseNet50-CS network, the model proposed in this paper improves by about 9.3% on the test set. Compared with the T3D algorithm, the model proposed in this paper improves by about 7.5% on the test set. Compared with ResNeXt-101, the model proposed in this paper improves by 0.88% on the test set. Compared with the Improved DenseNet algorithm, the model proposed in this paper improves by 1.48% on the test set. Experimental results show that the improved network in this paper can achieve good recognition effect on the HMDB51 dataset.
[0086] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.
Claims
1. A human motion recognition method, characterized in that The steps include: Step 1: Acquire multiple videos containing human body movements, and perform sparse sampling on each video to obtain multiple sparsely sampled video frame sequences; Step 2: Multiple video frame sequences are combined into a data set, and the data set is constructed into a training set and a test set; wherein each training sample in the training set includes a video frame sequence and a human action label corresponding to the video frame sequence; Step 3: Construct an action recognition model and use the training set to train the constructed action recognition model to obtain a trained action recognition model. The specific structure of the action recognition model is: A GRU timing processing module includes a 3D convolutional neural network and k bidirectional single-layer GRU units connected to the 3D convolutional neural network. The input end of the 3D convolutional neural network serves as the input end of the GRU timing processing module, and the output ends of the k bidirectional single-layer GRU units serve as the output end of the GRU timing processing module. The first spatial attention module has its input connected to the output of the GRU temporal processing module; The first residual module, whose input is connected to the output of the first spatial attention module; The second spatial attention module has its input connected to the output of the first residual module; The second residual module has its input connected to the output of the second spatial attention module; … The input of the Nth spatial attention module is connected to the output of the N-1th residual module; N is a positive integer; The input of the Nth residual module is connected to the output of the Nth spatial attention module; The multi-level feature fusion module includes N-1 feature fusion modules, namely the first feature fusion module, the second feature fusion module...the N-1th feature fusion module, the output end of the N-th residual module is connected to the input end of the first feature fusion module, the output end of the N-1th residual module is connected to the output end of the first feature fusion module and then connected to the input end of the second feature fusion module; the output end of the N-2th residual module is connected to the output end of the second feature fusion module and then connected to the input end of the third feature fusion module...the output end of the first residual module and the output end of the N-1th feature fusion module are connected and then serve as the output end of the multi-level feature fusion module; Segmented channel attention module, whose input is connected to the multi-level feature fusion module; A feature extraction layer, comprising a pooling layer connected to the output of the segmented channel attention module and a fully connected layer connected to the pooling layer, wherein the fully connected layer is the output of the action recognition model; Each of the above residual modules includes multiple identical residual blocks, and each residual block includes multiple convolutional layers connected in sequence. Each feature fusion module includes upsampling and convolution operations to make the dimension of its output end the same as the feature dimension of the output of its corresponding residual module, thereby enabling fusion. The specific processing steps of the segmented channel attention module are: The channel c corresponding to the feature map input to the segmented channel attention module is divided into M segments, and the number of channels in each of the M segments is obtained. The channel attention enhancement operation is performed on the M segments of the first feature map to obtain M segments of the second feature map. The M segments of the second feature map are concatenated and output along the channel dimension to obtain the output feature map of the segmented channel attention module. Step 4: Randomly select one of the samples in the test set and input it into the action recognition model trained in step 3 to obtain the human action recognition result.
2. The human motion recognition method according to claim 1, wherein: The 3D convolutional neural network in step 3 includes a convolution layer with 3 input channels, 64 output channels, a convolution kernel size of 7*7*7, a step size of (1, 2, 2), and a padding of (3, 3, 3).
3. The human motion recognition method according to claim 2, wherein: In step 3, between the 3D convolutional neural network and the k bidirectional single-layer GRU units, the following further comprises: a normalization layer and an activation function layer connected after the convolution layer in the 3D convolutional neural network.
4. The human motion recognition method according to any one of claims 1 to 3, characterized in that: The specific processing process of each spatial attention module in step 3 is as follows: Step 3-1: Perform dimension transformation on the image output by the GRU time series processing module to obtain the first image X. The dimension of X is (c, t, h, w), where c, t, h, and w represent the channel dimension, the number of frames, and the height and width of each frame of video, respectively. Step 3-2: compress the time dimension of the first image X to obtain a second image X', where the dimension of X' is transformed to (c, h, w); Step 3-3: After the second image X' passes through two convolutional layers conv1 and conv2 respectively, it is then activated by the sigmoid function to generate a corresponding weight mask for each position and output weighted. Finally, it is added to the first image X to obtain the final output F, which has the same dimension as X.
5. The human motion recognition method according to claim 4, wherein: The specific processing process of the channel attention enhancement operation in step 3 is: Step 3-a: Pool the spatiotemporal features except the channels in the M segments of the first feature map to compress their spatiotemporal dimensions to 1, thereby obtaining the first sub-map; Step 3-b: Perform a one-dimensional convolution operation on the channel dimension of the first sub-image to obtain the second sub-image; Step 3-c: Generate a corresponding weight mask for each channel position in the second sub-map through the activation function sigmoid operation and weighted output, and finally add the weighted output to the first feature map to obtain the second feature map.