A gait recognition model and a recognition method
By using a visual Transformer model and a parallel feature aggregation module, the accuracy and robustness issues of gait recognition in complex scenarios and under disguise conditions are solved, achieving efficient and accurate gait recognition.
Patent Information
- Application Number
- CN202510252911.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-05
- Publication Date
- 2025-11-07
- Estimated Expiration
- 2045-03-05
AI Technical Summary
Existing gait recognition methods suffer from performance degradation under varying perspectives, complex lighting conditions, and occlusion. Individual clothing variations lead to inaccurate gait feature extraction, and some deep learning methods require contour-based recognition, resulting in high computational resource consumption.
A visual Transformer model is used to extract general visual features. A parallel feature aggregation module extracts gait appearance features and dynamic denoising features respectively. A feature fusion and deepening module is used to weight and fuse and expand the feature dimensions. Finally, the most similar gait features are found in the database.
It improves the accuracy and robustness of gait recognition in complex scenarios and under disguise conditions, significantly enhances the accuracy of recognition results, and reduces the consumption of computing resources.
Smart Images
Figure CN120126216B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of gait recognition, and more particularly, to a gait recognition model and a recognition method. BACKGROUND
[0002] As an important biometric recognition technology, gait recognition realizes identity recognition by analyzing the walking pattern of an individual, and has advantages such as non-contact, long distance and difficult to counterfeit. However, the existing gait recognition methods still face many challenges in practical application. For example, the recognition performance of traditional methods significantly decreases under variable viewing angles, complex lighting and occlusion conditions; individual clothing changes (such as changing coats, carrying items) can lead to inaccurate gait feature extraction; and some deep learning methods require contour-based recognition, which requires a large amount of contour labeling and computing resources, limiting their practical application.
[0003] In recent years, the Vision Transformer (ViT) model has shown strong feature extraction capability in the field of computer vision, but its application in gait recognition is still in the exploratory stage. In addition, existing methods mostly use single-scale feature extraction strategies, which are difficult to capture both local details and global dynamic information of gait; and it is difficult to obtain accurate gait recognition results in complex scenes with variable dressing interference. SUMMARY
[0004] To overcome the defects of the prior art that contour labeling is required and the accuracy of gait recognition results is not high, the present application provides a gait recognition model and a recognition method without contour labeling and with high accuracy of recognition results.
[0005] To solve the above technical problems, the technical solutions of the present application are as follows:
[0006] A gait recognition model, comprising: a visual feature extraction and transformation module, a parallel feature aggregation module, a feature fusion deepening module and an output module;
[0007] The visual feature extraction and transformation module is used to extract general visual features from the input gait image of the visual feature extraction and transformation module and transform them into gait representations, the gait representations including gait appearance features and gait dynamic denoising features;
[0008] The parallel feature aggregation module is used to perform parallel feature extraction and weighted aggregation on the gait appearance features and the gait dynamic denoising features respectively, to obtain weighted aggregated gait appearance features and weighted aggregated gait dynamic denoising features;
[0009] The feature fusion deepening module is used for weighting and fusing the weighted aggregated gait appearance features and the weighted aggregated gait dynamic denoising features, and further deepening and expanding the feature dimension to refined features.
[0010] The output module is used for searching, from a preset database, gait features most similar to the refined features, and outputting the gait features most similar.
[0011] The application further provides a gait recognition model recognition method, comprising the following steps:
[0012] An image of a gait to be recognized is acquired.
[0013] The image of the gait to be recognized is input into the gait recognition model.
[0014] The gait recognition model outputs gait features most similar to the image of the gait to be recognized in a preset database, and obtains a recognition result of the image of the gait to be recognized.
[0015] Compared with the prior art, the technical scheme of the application has the beneficial effects that:
[0016] The application can effectively cope with gait recognition tasks under complex scenes and variable dressing conditions through the visual feature extraction and conversion module and the parallel feature aggregation mechanism, can fully capture individual gait features by combining local details and global dynamic information, can significantly improve recognition accuracy, and has strong feature expression capability; the feature fusion deepening module enhances the robustness of the model to noise and interference, and can obtain a gait recognition result with high accuracy. BRIEF DESCRIPTION OF DRAWINGS
[0017] Figure 1 A first structure diagram of the gait recognition model described in Embodiment 1 is shown.
[0018] Figure 2 A second structure diagram of the gait recognition model described in Embodiment 1 is shown.
[0019] Figure 3 A flow diagram of the gait recognition model recognition method described in Embodiment 2 is shown. DETAILED DESCRIPTION
[0020] The drawings are only used for illustrative purposes and cannot be understood as limiting the patent;
[0021] In order to better illustrate the embodiments, some components in the drawings may be omitted, enlarged or reduced, and do not represent the actual size of the product;
[0022] For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.
[0023] The technical solutions of the present application will be further described below with reference to the drawings and examples.
[0024] Example 1
[0025] The gait recognition model proposed in this example, Figure 1 The first structure diagram of the gait recognition model proposed in this example is shown in the figure. Figure 2 The second structure diagram of the gait recognition model proposed in this example is shown in the figure.
[0026] As shown in Figure 1 and Figure 2 The gait recognition model of this example includes a visual feature extraction and conversion module, a parallel feature aggregation module, a feature fusion deepening module, and an output module.
[0027] The visual feature extraction and conversion module is used to extract general visual features from the gait image input into the visual feature extraction and conversion module and convert them into gait representations, which include gait appearance features and gait dynamic denoising features.
[0028] The parallel feature aggregation module is used to perform parallel feature extraction and weighted aggregation on the gait appearance features and the gait dynamic denoising features respectively, to obtain weighted aggregated gait appearance features and weighted aggregated gait dynamic denoising features.
[0029] The feature fusion deepening module is used to weight fuse the weighted aggregated gait appearance features and the weighted aggregated gait dynamic denoising features and further deepen and expand the feature dimension to refined features.
[0030] The output module is used to find the most similar gait features from a pre-set database and output the most similar gait features.
[0031] In the specific implementation process, the visual feature extraction and conversion module and the parallel feature aggregation mechanism can effectively deal with gait recognition tasks under complex scenes and variable dressing conditions; and by combining local details and global dynamic information, individual gait features can be captured comprehensively, the recognition accuracy can be significantly improved, and the feature expression ability is strong; through the feature fusion deepening module, the robustness of the model to noise and interference is enhanced, and a gait recognition result with high accuracy can be obtained.
[0032] In an optional embodiment, a pre-trained Vision Transformer model and a gait representation extractor are sequentially connected in the visual feature extraction and conversion module, the visual feature extraction module extracts visual general features from the gait image input into the gait recognition model based on the pre-trained Vision Transformer model, and the gait representation extractor converts the visual general features into gait representations.
[0033] In an optional embodiment, the parallel feature aggregation module comprises at least a first sub-module and a second sub-module in parallel, wherein the first sub-module is configured to receive the gait appearance feature, the second sub-module is configured to receive the gait dynamic denoising feature, and each sub-module comprises at least two branches, a first feature concatenation layer, a first convolutional layer, a second convolutional layer and a first activation function layer;
[0034] The two branches comprise at least a first convolutional unit, a second convolutional unit and a third convolutional unit, wherein the first branch comprises at least the first convolutional unit, and the second branch comprises at least the second convolutional unit and the third convolutional unit connected in sequence;
[0035] The first convolutional unit comprises at least a third convolutional layer, a first batch normalization layer and a second activation function layer connected in sequence, the second convolutional unit comprises at least a fourth convolutional layer, a second batch normalization layer and a third activation function layer connected in sequence, and the third convolutional unit comprises at least a fifth convolutional layer, a third batch normalization layer and a fourth activation function layer connected in sequence;
[0036] The input end of the first branch, the input end of the second branch and the input end of the first convolutional layer are connected to the output end of the parallel feature aggregation module, the output of the first branch and the output of the second branch are input into the second convolutional layer after feature concatenation, and the output of the second convolutional layer and the output of the first convolutional layer are input into the first activation function layer after pixel-by-pixel addition.
[0037] In an optional embodiment, the feature fusion deepening module comprises an attention fusion unit and at least three residual network layers, wherein the input end of the first residual network layer and the input end of the second residual network layer are connected to the output end of the parallel feature aggregation module, the output end of the first residual network layer and the output end of the second residual network layer are connected to the input end of the attention fusion unit, the output end of the attention fusion unit is connected to the input end of the third residual network layer, the remaining residual network layers are connected in sequence, and the output end of the last residual network layer is the output end of the feature fusion deepening module;
[0038] The attention fusion unit comprises at least a second feature concatenation layer, a fourth convolutional unit, a fifth convolutional unit, a sixth convolutional layer and a Softmax layer connected in sequence, the attention fusion unit performs attention mechanism-based weighted fusion operation on the output of the Softmax layer to obtain the output result of the attention fusion unit;
[0039] The fourth convolution unit at least includes a seventh convolution layer, a fourth batch normalization layer and a fifth activation function layer connected in sequence, and the fifth convolution unit at least includes an eighth convolution layer, a fifth batch normalization layer and a sixth activation function layer connected in sequence.
[0040] As an example, as shown in Figure 2 , Figure 2 The numbers in the brackets represent the dimension of the features output by the convolution layer. The first convolution layer, the second convolution layer, the third convolution layer, the fourth convolution layer and the sixth convolution layer are all 1x1 convolution layers, the fifth convolution layer is a 3x3 convolution layer, the activation function of the first activation function layer is PReLU, and the activation functions in the second to sixth activation function layers are ReLU.
[0041] In an optional embodiment, the output unit at least includes a time pooling layer, a horizontal pyramid pooling layer, a full connection layer, a normalized feature separation layer and a loss function layer connected in sequence.
[0042] In an optional embodiment, a loss function Loss is arranged in the loss function layer, and an expression of the loss function Loss includes:
[0043] Loss = L tri + L ce + L gre
[0044] In the formula, L tri represents a three-state loss function, L ce represents a cross-entropy loss function, and L gre represents a loss generated by the gait representation extractor.
[0045] Embodiment 2:
[0046] The embodiment provides a recognition method of the gait recognition model in embodiment 1. Figure 3 A flowchart of the recognition method of the gait recognition model provided in the embodiment is shown in
[0047] As shown in Figure 3 , the recognition method of the gait recognition model includes the following steps:
[0048] S1: obtaining a gait image to be recognized;
[0049] S2: inputting the gait image to be recognized into the gait recognition model;
[0050] S3: the gait recognition model outputs a gait feature most similar to the gait image to be recognized in a preset database, and obtains a recognition result of the gait image to be recognized.
[0051] As an exemplary illustration, in actual application, the identity of the gait image to be identified can be determined as the identity corresponding to the most similar gait feature.
[0052] In an optional embodiment, before inputting the gait image to be identified into the gait recognition model, a training set including L consecutive frame gait images of a plurality of persons under a plurality of viewing angles and a plurality of gait conditions is obtained, where L is a preset positive integer.
[0053] The training set is input into the gait recognition model, and in the training process, the loss function is iteratively solved, and when the number of iterations reaches a preset value or a preset stop condition is met, the training is ended, and a trained gait recognition model is obtained.
[0054] When the gait image to be identified is input into the gait recognition model, the gait image to be identified is input into the trained gait recognition model.
[0055] As an exemplary illustration, when the loss function reaches a minimum value, i.e., a preset stop condition is met.
[0056] In an optional embodiment, the gait images in the training set are preprocessed gait images, and the step of preprocessing the gait images in the training set includes: performing a random horizontal flip operation on the gait images in the training set, and performing standardization processing on the gait images after the random horizontal flip, and inputting the gait images into the gait recognition model.
[0057] The embodiment also proposes a computer device including a memory and a processor, the memory stores computer readable instructions, and the computer readable instructions are executed by the processor to make the processor execute the steps of the gait recognition method according to the embodiment.
[0058] Embodiment 3:
[0059] Based on the gait recognition model according to embodiment 1 and the recognition method of the gait recognition model according to embodiment 2, the embodiment proposes a specific implementation example:
[0060] In the specific implementation process of the gait recognition model, the following steps can be performed:
[0061] Step 1: Select a gait dataset containing multiple perspectives, complex variable dressing conditions, and realistic shooting scenes to obtain high-quality RGB gait data. Standardize and enhance the data (e.g., random cropping, rotation, and illumination adjustment). Step 2: Input the preprocessed RGB gait data into a pre-trained ViT model to extract general gait features using its powerful feature extraction capabilities. By using a pre-trained model, error accumulation from upstream tasks (such as background segmentation and pose estimation) is avoided, and the computational cost of model training is significantly reduced. Then, a specialized gait feature extractor is used to convert the general features extracted by ViT into more targeted gait feature representations. Step 3: Design a parallel feature extraction module to extract local features of different body parts (such as leg swing and arm posture) and global features of the overall posture through multiple branches. Then, the features extracted by the feature extraction module are aggregated with weighted. By using a learnable weight distribution strategy, the contribution of each input channel is dynamically adjusted to ensure that important features are highlighted. At the same time, a residual connection structure (first convolutional layer) is used to preserve the original feature information and avoid feature loss. Finally, the optimized gait features are output. Step 4: Input the aggregated gait features into the feature fusion deepening module to extract more discriminative high-dimensional feature representations through multiple layers of nonlinear transformation (residual network layers). Then, use global pooling to reduce the dimension of the features, and use a fully connected layer for classification prediction (assign the gait image to be identified to the most similar identity, i.e., find the gait feature closest to the gait image to be identified). During training, use the triplet loss and cross-entropy loss functions for optimization, and continuously adjust the model parameters through the backpropagation algorithm until the model converges.
[0062] The triplet loss function is used to adjust the model's weights so that gait features belonging to the same person are closer, and gait features belonging to different people are farther apart. The cross-entropy loss function is used to increase the inter-class distance, and the loss function corresponding to the gait representation extractor is added to enable the gait representation extractor to convert general visual features into more targeted gait feature representations, allowing the model to accurately find the gait feature closest to the gait image to be identified.
[0063] In the specific implementation process of the gait recognition model recognition method, the gait recognition model goes through the following process:
[0064] Obtain the gait image to be identified: First, obtain the gait image to be identified. This data is usually represented by a dynamic sequence of human gait captured by a camera. The image can be a gait data composed of multiple frames from different angles and different times.
[0065] Extracting gait features: The gait recognition model is used to extract the features of the gait to be recognized, and the parallel feature aggregation gait recognition method based on the ViT model is used to effectively extract the high-dimensional representation reflecting the individual gait features from the dynamic sequence. These gait features can include motion information of various parts of the body, such as leg swing, stride change, etc.
[0066] Feature matching and similarity calculation: During the training process, the output module iteratively trains the gait recognition model through a pre-set loss function, so that the gait recognition model can match the gait features to be recognized with the most similar gait features in the pre-set database.
[0067] Dataset: (1) CASIA-B*, published by the Chinese Academy of Sciences Institute of Automation, consisting of 13,640 gait sequences from 124 people. This dataset includes three typical gait conditions: normal walking (NM), carrying a backpack (BG), and changing clothes (CL). In addition, it covers 11 different viewing angles, ranging from 0° to 180°, making it suitable for cross-view and cross-condition gait recognition tasks. It is considered one of the classic datasets in the field of gait recognition. The first 4 normal walking sequences (NM1-4) of each subject are used to construct the reference library (pre-set database), and NM5-6, BG1-2, CL1-2 are used as test data. (2) SUSTech1k published by the Southern University of Science and Technology, is a large-scale gait dataset, which includes 25,239 gait sequences from 1,000 people. This dataset includes various gait conditions and environmental changes, such as different clothes, changes in walking speed, and indoor / outdoor lighting changes, fully simulating the complexity of real-world scenarios. During the experiment, the first normal walking sequence (00-nm) of normal walking is used as the reference library (pre-set database), and 01-nm, bag (backpack), clothing (clothing), carrying (carrying), umbrella (umbrella), uniform (uniform), occlusion (occlusion), night (night), overall (overall) are used as test data.
[0068] Data preprocessing: The RGB format of L consecutive frames of gait images is used as input, which results in different horizontal displacements of L stacked frames due to different gait speeds and positions of each subject in the scene. Therefore, random horizontal flipping is used to adapt to different directions of gait patterns before inputting the data. Then the data is standardized and the image size is adjusted to fit the input format of the model while preserving key information.
[0069] Pre-trained ViT: Pre-trained on LVD-142M, it extracts general visual features suitable for multiple tasks through self-supervised learning. During training, part of the model's parameters are fine-tuned to adapt to the gait recognition task, while most of the parameters are still frozen to maintain the general ability of feature extraction and reduce computational cost. By extracting general features from pre-trained ViT, it avoids the error accumulation caused by using upstream tasks to extract gait information such as contours and skeletons in other gait recognition methods. As shown in Figure 2
[0070] BigGait: Learning Gait Representation You Want by Large Vision ModelsDingqiang Ye, Southern University of Science and Technology): The general features are converted into gait representations through three different branches. The Mask branch generates a gait foreground mask, effectively eliminating background noise at the global level. The Appearance branch captures and processes the preserved foreground features, extracting gait-related visual information such as contours and static attributes. To further reduce the impact of high-frequency noise, such as clothing texture, the Denoising branch uses smoothing operations in parallel with the Appearance branch to reduce noise interference. Finally, the Parsing-like representation and Denoising representation are generated as inputs to the PFA module.
[0071] Parallel Feature Aggregation (PFA) module: (1) Firstly, gait features are extracted by two branches collaboratively: Branch 1 adopts a 1x1 convolutional layer to extract basic features while reducing computational overhead. Branch 2 combines 1x1 and 3x3 convolutional layers to capture precise local gait information. Compared with serially stacking multiple convolutional layers, this parallel structure can model features more efficiently. (2) Gait information from both branches is concatenated and a 1x1 convolutional operation is performed along the channel dimension by learning the weights, which measures the contribution of each input channel and assigns different weights to generate a new output feature map. This step effectively removes redundant functions while emphasizing key information. Compared with larger convolutional kernels, 1x1 convolution does not require complex calculations along the spatial dimension, thus enabling effective channel information extraction while reducing the computational burden of the model. (3) Subsequently, residual connections are used to preserve input features and ensure stable gradient flow during deep model training. Finally, the PRELU activation function introduces other nonlinear features, which, with its adaptive learning ability, enhances the model's ability to recognize complex dynamic gait changes, thereby improving the model's robustness and accuracy in challenging situations. Parsing-like representation and denoising representation extract gait features through parallel branches of the PFA module and assign different weights to effectively integrate appearance and dynamic information.
[0072] Subsequently, the two gait representations are input into the Resnet module to further optimize the information from different branches and achieve weighted integration through the attention fusion module, and then gradually expand the dimension of gait features through multiple Resnet modules to capture more complex gait patterns. Finally, we use a widely used framework to identify the extracted gait functions. In the temporal pooling layer, temporal information in the gait sequence is aggregated, and then horizontal pyramid pooling is performed for horizontal cutting to extract features from different parts of the gait sequence, respectively. These cut blocks are then input into separate fully connected layers to generate new features, which are then passed to the normalized feature separation layer for normalization and classification, and finally output the loss.
[0073] Loss function: Triplet loss is used to construct sample pairs and optimize the distribution of features in the embedding space. It ensures that gait features of the same person are closer, while gait features of different people are increasingly distant, thereby reducing intra-class distance. At the same time, cross-entropy loss is used to increase inter-class distance, thereby improving the classification performance of the model. The calculation of the loss Loss can be represented by the following formula:
[0074] Loss = L tri + L ce + L gre
[0075] where L tri , L ce , Lgre respectively represent the ternary loss, the cross-entropy loss, and the loss generated by the GRE module (gait representation extractor). In order to evaluate the performance of different methods, we use the Rank-1 precision metric as the evaluation standard.
[0076] The high-confidence VIT-enhanced parallel feature aggregation network not only can perform end-to-end gait recognition, but more importantly, effectively captures and focuses on key gait features by using the PFA module, thereby achieving stronger gait recognition.
[0077] The recognition method of the gait recognition model is a parallel feature aggregation gait recognition method based on a ViT model. Compared with traditional gait recognition methods, the main features of the method are as follows: (1) a pre-trained visual Transformer model is introduced to avoid error accumulation and reduce training costs. Traditional methods usually need complex upstream task processing (such as background segmentation, pose estimation, etc.), which is easy to cause error accumulation. The present application directly uses a ViT model pre-trained on a large amount of data to extract features, thereby avoiding the error transmission problem caused by upstream task processing and reducing the computational cost of model training. (2) a parallel feature aggregation module is proposed. Traditional methods often use a single scale of features, which is difficult to fully capture the local details and global dynamic information of the gait. The present application designs a parallel feature aggregation module, which simultaneously extracts local features of individual body parts and global features of the overall posture through multiple branches, and combines short-term motion patterns (such as single-step swing) and long-term motion patterns (dynamic changes in the entire gait cycle), thereby more comprehensively describing the gait features of individuals. Then, through an adaptive feature aggregation module, a learnable weight distribution strategy is used to dynamically adjust the feature contribution of each input channel, thereby optimizing the feature representation. At the same time, a residual connection structure is used to retain the original feature information and avoid feature degradation problems. (3) stronger robustness and generalization ability. The method of the present application shows stronger robustness and generalization ability in complex scenes (such as multi-view, variable dressing conditions, and real shooting environments). By combining pre-trained ViT and feature extraction, it can effectively cope with challenges such as light changes, occlusions, and clothing changes, thereby significantly improving the accuracy of gait recognition.
[0078] In summary, the advantages of the present application are as follows: strong adaptability: through the pre-trained ViT model and the parallel feature aggregation mechanism, the gait recognition task in complex scenes and variable dressing conditions can be effectively coped with; strong feature expression ability: by combining local details and global dynamic information, individual gait features are fully captured, thereby significantly improving recognition accuracy; good robustness: through adaptive feature aggregation and residual connection, the robustness of the model to noise and interference is enhanced.
[0079] The present application can be widely applied in the fields of security monitoring, intelligent transportation, medical rehabilitation, etc., and provides reliable technical support for individual identity recognition and behavior analysis.
[0080] The present embodiment also proposes Table 1 and Table 2 for showing the recognition accuracy of the gait recognition model, which are shown as follows.
[0081] Table 1 shows the recognition accuracy (%) of different views in the dataset. CASIA-B * Table 2 shows the recognition accuracy (%) of different scenes in the SUSTech1k dataset.
[0082]
[0083] Table 2 shows the recognition accuracy (%) of different scenes in the SUSTech1k dataset.
[0084] deepgait ]]> deepgait ]]> gaitbase ]]> BigGai t]]> Ours Normal 54.63 82.71 77.29 94.1 95.47 Bag 57.08 80.53 75.82 94.65 96.03 Clothing 25.08 48 45.43 67.13 69.57 Carrying 53.42 78.12 73.31 94.54 96.01 Umberalla 52.35 80.58 71.84 93.85 95.44 Uniform 44.68 79.29 72.2 88.07 91.44 Occlusion 49.85 85.41 78.62 99.03 99.68 Night 21.05 25 24.09 75.85 82.19 Overall 53.05 78.61 73.39 93.63 95.09
[0085] The same or similar reference numerals refer to the same or similar components;
[0086] The terms describing the positional relationship in the drawings are only used for illustrative description, and should not be understood as a limitation on the present patent;
[0087] Obviously, the above embodiments of the present application are only examples for clearly illustrating the present application, and are not intended to limit the embodiments of the present application. Based on the above description, other different forms of changes or variations can be made by those skilled in the art. Here, it is not necessary and impossible to enumerate all the embodiments. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application should be included in the protection scope of the claims of the present application.
Claims
1. A gait recognition model, characterized in that, The application relates to a visual feature extraction and conversion module, a parallel feature aggregation module, a feature fusion deepening module and an output module. The visual feature extraction and conversion module is used for extracting general visual features from gait images of an input visual feature extraction and conversion module and converting the general visual features into gait representations, wherein the gait representations include gait appearance features and gait dynamic denoising features. The parallel feature aggregation module is used for performing parallel feature extraction and weighted aggregation on the gait appearance features and the gait dynamic denoising features respectively, so as to obtain weighted aggregated gait appearance features and weighted aggregated gait dynamic denoising features. The feature fusion deepening module is used for weightedly fusing the weighted aggregated gait appearance features and the weighted aggregated gait dynamic denoising features and further deepening and expanding the feature dimensions into refined features. The output module is used for searching, from a preset database, gait features most similar to the refined features and outputting the most similar gait features. The parallel feature aggregation module at least includes parallel first and second submodules, wherein the first submodule is used for receiving the gait appearance features, the second submodule is used for receiving the gait dynamic denoising features, and each submodule at least includes two branches, a first feature splicing layer, a first convolutional layer, a second convolutional layer and a first activation function layer. The two branches at least include first, second and third convolutional units, wherein the first branch at least includes the first convolutional unit, and the second branch at least includes the second and third convolutional units connected in sequence. The first convolutional unit at least includes third, first batch normalization and second activation function layers connected in sequence, the second convolutional unit at least includes fourth, second batch normalization and third activation function layers connected in sequence, and the third convolutional unit at least includes fifth, third batch normalization and fourth activation function layers connected in sequence. The input ends of the first and second branches and the input end of the first convolutional layer are connected with the output end of the parallel feature aggregation module, the output of the first branch and the output of the second branch are spliced and input into the second convolutional layer, and the output of the second convolutional layer and the output of the first convolutional layer are added pixel by pixel and input into the first activation function layer. A pre-trained Vision Transformer model and a gait representation extractor are sequentially connected in the visual feature extraction and conversion module, the visual feature extraction module extracts visual general features from gait images input into a gait recognition model based on the pre-trained Vision Transformer model, and the gait representation extractor converts the visual general features into gait representations.
2. The gait recognition model of claim 1, wherein, 3. The gait recognition model of claim 1, wherein, The feature fusion deepening module comprises an attention fusion unit and at least three residual network layers, wherein the input ends of the first and second residual network layers are connected to the output ends of the parallel feature aggregation modules, the output ends of the first and second residual network layers are connected to the input ends of the attention fusion unit, the output end of the attention fusion unit is connected to the input end of the third residual network layer, and the remaining residual network layers are connected in sequence, and the output end of the last residual network layer is the output end of the feature fusion deepening module; The attention fusion unit comprises at least a second feature splicing layer, a fourth convolution unit, a fifth convolution unit, a sixth convolution layer and a Softmax layer connected in sequence, and the attention fusion unit performs attention mechanism-based weighted fusion operation on the output of the Softmax layer to obtain the output result of the attention fusion unit. The fourth convolution unit comprises at least a seventh convolution layer, a fourth batch normalization layer and a fifth activation function layer connected in sequence, and the fifth convolution unit comprises at least an eighth convolution layer, a fifth batch normalization layer and a sixth activation function layer connected in sequence.
4. The gait recognition model of any one of claims 1-3, wherein, The output unit comprises at least a temporal pooling layer, a horizontal pyramid pooling layer, a full connection layer, a normalized feature separation layer and a loss function layer connected in sequence.
5. The gait recognition model of claim 4, wherein, The loss function layer is internally provided with a loss function The expression of the loss function includes: wherein denotes a ternary loss function, denotes a cross-entropy loss function, denotes a loss produced by the gait representation extractor.
6. A gait recognition model recognition method according to claim 5, characterized in that, The method comprises the following steps: obtaining a gait image to be identified; inputting the gait image to be identified into the gait recognition model; the gait recognition model outputs the gait feature most similar to the gait image to be identified in the preset database, and obtains the recognition result of the gait image to be identified.
7. The gait recognition method according to claim 6, characterized in that, Before inputting the gait image to be identified into the gait recognition model, a training set comprising L frames of continuous frame gait images of a plurality of persons under a plurality of viewing angles and a plurality of gait conditions is obtained, wherein L is a preset positive integer; inputting the training set into the gait recognition model, iteratively solving the loss function in the training process, ending the training when the number of iterations reaches a preset value or meets a preset stopping condition, and obtaining the trained gait recognition model; when the gait image to be identified is input into the gait recognition model, the gait image to be identified is input into the trained gait recognition model.
8. The gait recognition method of claim 7, wherein, The gait images in the training set are preprocessed gait images, and the step of preprocessing the gait images in the training set comprises: performing random horizontal flip operation on the gait images in the training set, and performing standardization processing on the gait images after random horizontal flip, and inputting the gait images into the gait recognition model. 9.A computer device, comprising a memory and a processor, wherein the memory stores computer readable instructions, and the computer readable instructions are configured to cause the processor to perform the method according to any one of claims 1 to 8. The computer readable instructions are executed by the processor to make the processor execute the steps of the gait recognition method according to any one of claims 6-8. The computer readable instructions are executed by the processor to make the processor execute the steps of the gait recognition method according to any one of claims 6-8.