Gait recognition method, system, medium, device and product

Abstract

The present application relates to the technical field of image processing, and discloses a gait recognition method, which captures appearance contour information from different angle positions by polar coordinate sampling at R different sampling angles, explicitly models contour structure in a polar coordinate system, makes up for the deficiency of traditional appearance-based methods in capturing structural features, and thus improves the robustness of gait recognition to appearance changes. Furthermore, the present application also performs multi-scale fusion on structural features of different sampling angles to suppress redundant information and enhance the stability and distinguishability of structural representation. In addition, the present application combines gait features with global information and local information on the final structural features obtained by polar coordinate sampling, maintains the robustness to appearance disturbance, and improves the sensitivity of the model to shape changes. The present application effectively alleviates the influence of view angle change, clothing change and carrying articles, and greatly improves the robustness of appearance contour modeling.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a gait recognition method, system, medium device, and product. Background Technology

[0002] Gait recognition, with its non-contact data acquisition capabilities, has shown broad application potential in fields such as disease diagnosis, emotion analysis, and criminal investigation. However, in real-world applications, gait recognition systems still face considerable challenges due to the complexity of data acquisition, such as different camera angles, diverse walking conditions, and unexpected occlusions.

[0003] Gait recognition methods are generally divided into two categories: model-based methods and appearance-based methods. Model-based methods rely on explicit human structural information, capturing posture and kinematic changes during walking by modeling skeletal key points, joint topology, or parametric human models. Appearance-based gait recognition methods mainly utilize contour sequences during walking, extracting discriminative identity features by analyzing the overall contour and local limb movements.

[0004] In recent years, deep learning has significantly propelled the development of appearance-based gait recognition technology. Most existing methods learn discriminative representations from contour sequences by employing global or local feature modeling strategies. Global methods emphasize extracting overall gait patterns from the entire contour, while local methods aim to capture fine motion features of individual body parts.

[0005] However, the aforementioned global or local feature modeling strategies all heavily rely on the stability and consistency of the appearance contour. In real-world applications, due to factors such as changes in camera angle, clothing occlusion, and fluctuations in lighting conditions, pedestrian appearance contours exhibit significant differences under different viewpoints. Especially at side, oblique, or pitch angles, the contour shape may suffer severe distortion or information loss. This viewpoint sensitivity makes it difficult for existing appearance-based methods to construct robust and generalizable appearance representation models, resulting in insufficient structural modeling capabilities for appearance contours. This limits their recognition performance and applicability in complex real-world environments. Summary of the Invention

[0006] The purpose of this invention is to provide a more robust gait recognition method, system, medium, device, and product.

[0007] To achieve the above objectives, the present invention provides the following solution: In a first aspect, the present invention provides a gait recognition method, comprising: Define R different sampling angles, determine the center point of each sampling angle, where R is an integer greater than 1; at each sampling angle, take the corresponding center point as the origin of polar coordinates and sample radially to obtain human body structure data, and extract the structural features of the sampling angle based on the human body structure data; Gait recognition is performed based on the structural features at the R different sampling angles.

[0008] In a second aspect, the present invention provides a gait recognition system, comprising: The structure perception module is used to define R different sampling angles, determine the center point of each sampling angle, where R is an integer greater than 1; at each sampling angle, sampling is performed radially with the corresponding center point as the origin of polar coordinates to obtain human body structure data, and the structural features of the sampling angle are extracted based on the human body structure data. The recognition module is used to perform gait recognition based on the structural features at the R different sampling angles.

[0009] Thirdly, the present invention provides an electronic device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the gait recognition method described above.

[0010] Fourthly, the present invention provides a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the gait recognition method described above.

[0011] Fifthly, the present invention provides a computer program product comprising a computer program that, when executed by a processor, implements the steps of the gait recognition method described above.

[0012] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention captures appearance contour information from different angular positions by performing polar coordinate sampling at R different sampling angles. It explicitly models the contour structure in a polar coordinate system, overcoming the shortcomings of traditional appearance-based methods in capturing structural features, thereby improving the robustness of gait recognition to appearance changes. Furthermore, this invention performs multi-scale fusion of structural features from different sampling angles to suppress redundant information and enhance the stability and distinguishability of the structural representation. In addition, this invention combines gait features with both global and local information into the final structural features obtained from polar coordinate sampling, improving the model's sensitivity to shape changes while maintaining robustness to appearance perturbations. This invention effectively mitigates the effects of changes in viewpoint, clothing, and carried items, significantly improving the robustness of appearance contour modeling. Attached Figure Description

[0013] Figure 1 This is a schematic diagram illustrating the acquisition of the final structural features according to an embodiment of the present invention; Figure 2 This is a flowchart of the gait recognition system according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the gait recognition method according to an embodiment of the present invention; Figure 4 This is a schematic diagram simulating the effect of different occlusion regions on gait sequences in an occlusion experiment; Figure 5 This is a diagram showing how the recognition accuracy changes as the size of the occluded area increases; Figure 6 This is a schematic diagram of the device structure of the hardware operating environment involved in the gait recognition method in this embodiment of the invention. Detailed Implementation

[0014] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0015] Example 1 like Figures 1 to 3 As shown, a preferred embodiment of the gait recognition method of the present invention includes: Define R different sampling angles, determine the center point of each sampling angle, where R is an integer greater than 1; at each sampling angle, take the corresponding center point as the origin of polar coordinates and sample radially to obtain human body structure data, and extract the structural features of the sampling angle based on the human body structure data; Gait recognition is performed based on the structural features at the R different sampling angles.

[0016] This embodiment captures appearance contour information from different angles by performing polar coordinate sampling at R different sampling angles, and explicitly models the contour structure in the polar coordinate system, which makes up for the shortcomings of traditional appearance-based methods in capturing structural features, thereby improving the robustness of gait recognition to appearance changes.

[0017] Specifically, in this embodiment, the center point of each sampling angle is the same point. Therefore, this embodiment takes a point inside or near the human body contour as the polar coordinate center point, and samples are taken radially around this center point at R different sampling angles. The radial distance of the human body contour boundary point relative to the center point at the sampling angle is determined based on the sampled data. The human body contour can then be modeled based on the sampling angle and the radial distance.

[0018] At each of the stated sampling angles, sampling is performed radially with the corresponding center point as the origin of polar coordinates to obtain human body structure data, including: At each sampling angle, N sampling points are set radially with the corresponding center point as the origin of polar coordinates, where N is an integer greater than 1; The center point and the N sampling points are sampled respectively to obtain N+1 structural point data, and the N+1 structural point data constitute the human body structure data; Feature extraction is performed on each of the N+1 structural point data to obtain N+1 structural point features; The structural features are generated by weighting and linearly combining the N+1 structural point features according to their respective weights.

[0019] In this embodiment, N+1 sampling points (including the center point + N sampling points) are set at each sampling angle, wherein the distance between any two adjacent sampling points is equal, and uniform sampling is performed.

[0020] Furthermore, in this embodiment, a sampling distance is assigned at each sampling angle, and the sampling distances at each sampling angle may be equal or unequal.

[0021] This embodiment defines R different sampling angles for the contour feature map and extracts the structural features of each sampling angle. Therefore, before defining the R different sampling angles, this embodiment obtains a gait image sequence to be identified and acquires several contour feature maps based on the gait image sequence. Specifically, this embodiment inputs the gait image sequence to be identified into the backbone network to obtain the contour feature maps. The backbone network is a deep neural network, such as a CNN or Transformer architecture, that extracts basic spatiotemporal features from the original image sequence.

[0022] like Figure 1As shown, in this embodiment, the gait image sequence to be identified is input into the backbone network, and the backbone network generates feature maps. Starting from C, perform a 1×1 convolution to project the channel dimension from C to R, resulting in... , It is a contour feature map, and the number of channels after projection of the backbone network corresponds to the number of sampling angles. Each channel after projection is associated with a predefined sampling angle.

[0023] For a given contour feature map, taking a point inside or near the human contour on the feature map as the polar coordinate center point, R sampling angles are defined around this center point, and a sampling distance is assigned to each sampling angle. For each sampling angle, N+1 sampling points (including the center point) are uniformly sampled, and these sampling points are adaptively fused to extract structural features. Specifically, the sampling distance is set to N pixels, corresponding to N uniformly spaced sampling intervals along each sampling angle.

[0024] For each spatial location (x, y), N+1 points are uniformly sampled along the r-th angle, including the center point. For example... Figure 1 As shown in part a of the diagram. The coordinates of the nth sampling point are defined as follows:

[0025]

[0026] Where n = 0 corresponds to the center pixel (x, y), that is , Sampling distance, This represents the number of sampling angles.

[0027] Based on the above sampling points, define the structural features at the r-th angle. for:

[0028] in, The structural point features of the nth sampling point at the rth angle; To prevent the suppression of informative structural points during the fusion process, learnable weights are assigned to the N+1 sampling points, with the weights normalized by the output of a 1×1 convolution and Softmax, such as... Figure 1 Part b is shown in the image:

[0029] in and This represents the weights and biases in the convolutional layer.

[0030] The structural features at the R sampling angles are calculated as follows:

[0031]

[0032] By integrating multi-angle ray sampling and adaptive point-by-point weighting, this embodiment transforms local contour features into a well-defined angular structural representation. This design preserves fine structural cues at different angular positions while reducing redundancy introduced by dense spatial sampling, thereby enabling more effective extraction of structural features of the human body contour.

[0033] In this embodiment, R=64 and N=5, that is, 64 sampling angles are set, and the number of sampling points at each sampling angle is set to 5.

[0034] Furthermore, the gait recognition based on the structural features from the R different sampling angles includes: The structural features from the R different sampling angles are fused to obtain the final structural features; Gait recognition is performed based on the final structural features.

[0035] This embodiment also fuses structural features from different sampling angles to suppress redundant information and enhance the stability and distinguishability of structural representation.

[0036] In polar coordinate representation, structural features extracted from different angular positions capture complementary shape information of the human body contour. Processing each angular feature independently can lead to fragmented representation and increased sensitivity to local perturbations. For example... Figure 1 As shown in part c, multi-scale convolution is used to aggregate structural features at adjacent angular positions, thereby achieving effective information integration and a more stable structural representation:

[0037] Where ConvK represents applying a convolution with kernel size K along the angular dimension, and normalizing the weights. The calculation is as follows:

[0038]

[0039] This embodiment effectively integrates complementary shape information from different angular positions by adaptively fusing angular structural features at multiple scales, thereby generating a more stable and discriminative structural representation.

[0040] This embodiment fuses structural features from different sampling angles by performing multi-scale fusion of structural features from adjacent sampling angles. For example, a sliding window with sizes of 3 and 5 is used. The input is R, and the output is still R, except that adjacent angles are fused together. Then, the outputs of the two sliding windows are adaptively fused. Specifically: R structural features from different sampling angles are received. The structural features from the R different sampling angles are traversed using a sliding window. For each current sampling angle structural feature, a first neighborhood window with a scale of 3 is extracted; a second neighborhood window with a scale of 5 is extracted; a first fusion operation is performed on the structural features within the first neighborhood window to obtain a first feature; a second fusion operation is performed on the structural features within the second neighborhood window to obtain a second feature; the first feature and the second feature are adaptively weighted and fused to obtain the output structural features for the current sampling angle; the output structural features from the R different sampling angles are combined into the final structural feature.

[0041] Furthermore, the gait recognition method in this embodiment also includes: Gait features are extracted, including global features representing the overall contour movement pattern of the human body and local features representing the movement changes of body parts; In the gait recognition based on the final structural features, gait recognition is performed according to the gait features and the final structural features.

[0042] This embodiment combines gait features with global and local information into the final structural features obtained by polar coordinate sampling, thereby improving the model's sensitivity to shape changes while maintaining robustness to appearance perturbations.

[0043] The step of gait recognition based on the gait features and the final structural features includes: The gait features are fused with the final structural features to obtain fused features; The fused features are sequentially processed by temporal pooling, horizontal pyramid pooling, a first fully connected layer and batch normalization, and a second fully connected layer to output the gait identity category.

[0044] In this embodiment, the gait features are first concatenated with the final structural features, and then fed into a convolutional layer for feature integration to obtain fused features.

[0045] Furthermore, the extraction of gait features includes global features representing the overall contour movement pattern of the human body and local features representing changes in the movement of body parts, including: A sequence of gait images to be identified is obtained, and features are extracted from the gait image sequence using a three-dimensional convolutional network to obtain spatiotemporal features. The gait features are extracted based on the spatiotemporal features. In this embodiment, the spatiotemporal features are input into a Global-Local Feature Extractor (GLFE) to extract the gait features.

[0046] Example 2 like Figure 2 As shown, the gait recognition method of this invention includes: S201. Obtain the gait image sequence to be identified; S202. The gait image sequence to be identified is subjected to feature extraction through a three-dimensional convolutional network to obtain spatiotemporal features; the spatiotemporal features are input into a global-local feature extractor to obtain gait features, the gait features include global features representing the overall contour movement pattern of the human body and local features representing the movement changes of body parts; S203. Input the gait image sequence to be identified into the backbone network to obtain a contour feature map; define R different sampling angles for the contour feature map, determine the center point of each sampling angle, where R is an integer greater than 1; at each sampling angle, take the corresponding center point as the origin of polar coordinates and sample radially to obtain human body structure data, and extract the structural features of the sampling angle based on the human body structure data. S204. The structural features from the R different sampling angles are fused to obtain the final structural features; S205. The gait features are fused with the final structural features to obtain fused features; S206. The fused features are sequentially processed by time pooling, horizontal pyramid pooling, first fully connected layer and batch normalization, and second fully connected layer to output the gait identity category.

[0047] Furthermore, in step S206, the output of the second fully connected layer is typically used to calculate the cross-entropy loss function and the triplet loss function, and the cross-entropy loss function and the triplet loss function are used in combination to improve the discriminative ability of the model used in the gait recognition method.

[0048] The other steps in this embodiment are the same as in Embodiment 1, and will not be repeated here.

[0049] Example 3 like Figure 3 As shown, an embodiment of the present invention provides a gait recognition system, including: The structure perception module 304 is used to define R different sampling angles, determine the center point of each sampling angle, where R is an integer greater than 1; at each sampling angle, sampling is performed radially with the corresponding center point as the origin of polar coordinates to obtain human body structure data, and the structural features of the sampling angle are extracted based on the human body structure data. The recognition module 307 is used to perform gait recognition based on the structural features at the R different sampling angles.

[0050] Specifically, the gait recognition system of this embodiment of the invention further includes: The spatiotemporal feature extraction module 301 is used to acquire the gait image sequence to be identified, and to extract features from the gait image sequence through a three-dimensional convolutional network to obtain spatiotemporal features; A global-local feature extractor 302 is used to extract the gait features based on the spatiotemporal features. The gait features include global features representing the overall contour movement pattern of the human body and local features representing the movement changes of body parts. The contour feature extraction module 303 is used to obtain contour feature maps based on the gait image sequence to be identified; the contour feature extraction module 303 adopts a backbone network; The structure perception module 304 is used to define R different sampling angles for the contour feature map, determine the center point of each sampling angle, where R is an integer greater than 1; at each sampling angle, sampling is performed radially with the corresponding center point as the origin of polar coordinates to obtain human body structure data, and structural features of the sampling angle are extracted based on the human body structure data; wherein, at each sampling angle, N sampling points are set radially with the corresponding center point as the origin of polar coordinates, where N is an integer greater than 1; sampling is performed on the center point and the N sampling points respectively to obtain N+1 structural point data, which constitute the human body structure data; feature extraction is performed on the N+1 structural point data respectively to obtain N+1 structural point features; and the N+1 structural point features are weighted linearly combined according to their respective weights to generate the structural features. The multi-scale structural feature fusion module 305 is used to fuse the structural features from the R different sampling angles to obtain the final structural features; in this embodiment, the multi-scale structural feature fusion module 305 adopts a multi-scale configuration combining convolution kernel sizes of 3 and 5.

[0051] The feature fusion module 306 is used to fuse the gait features with the final structural features to obtain fused features; in this embodiment, the feature fusion module 306 adopts a convolutional layer. The recognition module 307 is used to sequentially process the fused features through temporal pooling, horizontal pyramid pooling, a first fully connected layer and batch normalization, and a second fully connected layer, and output the gait identity category. In this embodiment, the recognition module 307 includes a temporal pooling layer, a horizontal pyramid pooling layer (HPP), a first fully connected layer and batch normalization layer (FC&BN), and a second fully connected layer (FC) connected sequentially.

[0052] In this embodiment, the gait recognition system is defined as GaitSF, which is a structural feature modeling framework for gait recognition. The architecture of GaitSF consists of two feature modeling branches. The global-local feature extraction branch (global-local feature extractor 302) aims to capture global appearance semantics as well as fine-grained local motion changes, while the structural feature extraction branch (structure perception module 304 and multi-scale structural feature fusion module 305) focuses on modeling the structural distribution of human body contours at different angles.

[0053] For global and local feature modeling, the global-local feature extraction branch follows the design concept of the global-local feature extractor proposed in GaitGL, which preserves the overall contour information while enhancing the discriminability of specific motion patterns of body parts.

[0054] The structural feature extraction branch consists of two key components: a structure perception module 304 and a multi-scale structural feature fusion module 305. The structure perception module 304 captures the structural features of the human body contour through multi-angle ray sampling and adaptive fusion, thereby enhancing sensitivity to local shape deformations. The multi-scale structural feature fusion module 305 further adaptively fuses structural features from adjacent angular positions to suppress redundant information and improve the stability and distinguishability of the structural representation.

[0055] Unlike traditional convolutions that operate within a fixed receptive field, the structure-aware module 304 explicitly extracts the structural features of the human body contour through multi-angle sampling in polar coordinates.

[0056] This embodiment proposes GaitSF, a structural feature modeling framework for gait recognition, which overcomes the limitations of traditional contour-based methods in explicitly capturing contour structures. By explicitly modeling contour structures in polar coordinates, this framework compensates for the shortcomings of traditional appearance-based methods in capturing structural features, thereby improving the robustness of gait recognition to appearance changes. Specifically, the structure-aware module 304 captures the structural features of the human contour through multi-angle sampling and adaptive fusion, while the multi-scale structural feature fusion module 305 further adaptively fuses structural features at different angular scales to suppress redundant information and enhance the stability and discriminability of the structural representation. These two modules work together to improve the model's sensitivity to shape changes while maintaining robustness to appearance perturbations. Extensive experiments on three benchmark datasets, including CASIA-B, SUSTech1K, and OU-MVLP, demonstrate that GaitSF outperforms state-of-the-art gait recognition methods. Furthermore, ablation experiments and transfer experiments validate the effectiveness and modularity of the proposed structural feature branch. Occlusion experiments show that the proposed method can maintain stable performance under square occlusion, indicating that it has strong occlusion robustness.

[0057] Dataset Comparison Experiment 1: CASIA-B Dataset On the CASIA-B dataset, we used Rank-1 accuracy as the performance evaluation metric and compared it with several representative gait recognition methods. The results are shown in Table 1. Overall, the proposed GaitSF method achieved the best average recognition performance under all three conditions, indicating that the model possesses good stability and generalization ability under various appearance interferences.

[0058] Specifically, under both Normal Condition (NM) and Backpack / Carrying Bag Condition (BG), this method significantly outperforms other methods in most perspectives, achieving average Rank-1 accuracies of 98.5% and 96.1%, respectively. Compared to recently high-performing methods such as GaitCTCG, GaitDAN, and SPOSGait, this method improves accuracy by 0.4%, 0.7%, and 1.0% under NM conditions, and by 2.1%, 0.9%, and 1.7% under BG conditions. Under the Clothing Change Condition (CL), despite slight fluctuations in performance from individual perspectives, this method still achieves an average accuracy of 87.2%, significantly outperforming GaitCTCG, GaitDAN, and SPOSGait, with improvements of 4.2%, 1.2%, and 3.0%, respectively. It is important to emphasize that the large-scale occlusion and deformation of body contours in CL scenes are usually the most challenging scenarios in gait recognition, and the stable performance of this method further verifies the effectiveness of structural features under conditions of appearance uncertainty.

[0059] Table 1. Performance evaluation metrics of various gait recognition methods on the CASIA-B dataset.

[0060] Dataset Comparison Experiment 2: OU-MVLP Dataset As shown in Table 2, we evaluated the generalization performance of GaitSF on the large-scale OU-MVLP dataset and compared it with several representative methods. Experimental results show that the proposed method maintains high recognition accuracy on this dataset, with average Rank-1 accuracy improvements of 0.5%, 0.7%, and 0.5% compared to GMSN, GaitDAN, and GaitSCM, respectively. It is worth noting that the OU-MVLP dataset is large in scale, covers a wide range of perspectives, and exhibits significant individual variability, thus placing higher demands on the model's cross-domain generalization ability and robustness. Under these challenging conditions, the proposed method demonstrates consistent performance, indicating that the adopted feature modeling strategy is not limited to medium-sized datasets but also maintains reliable discriminative ability on large-scale, multi-view gait datasets.

[0061] Table 2 Comparison of generalization performance of various gait recognition methods on the OU-MVLP dataset.

[0062] Dataset Comparison Experiment 3: SUSTeck1K Dataset Table 3 summarizes the performance of GaitSF on the SUSTech1K dataset under different environments. The proposed method achieves an overall Rank-1 accuracy of 77.9% on this dataset, indicating its effectiveness under various environmental conditions. This result improves upon DeepGaitV2 and GaitBase by 0.5% and 1.8%, respectively. Under clothing conditions (where changes in clothing lead to significant changes in appearance), the proposed method achieves a Rank-1 accuracy of 54.5%, improving upon DeepGaitV2 and GaitBase by 8.2% and 4.9%, respectively. These results demonstrate that explicitly modeling structural features helps improve the robustness of the silhouette to significant perturbations caused by clothing changes.

[0063] Table 3 Performance comparison of GaitSF on the SUSTech1K dataset under different environments

[0064] Furthermore, this embodiment also provides ablation experiments using the global-local feature extractor 302, the structure-aware module 304, and the multi-scale structural feature fusion module 305 individually or in combination. Specifically, GLFE is the global-local feature extractor 302, SPM is the structure-aware module 304, MSFF is the multi-scale structural feature fusion module 305, and SFE is the combination of the structure-aware module 304 and the multi-scale structural feature fusion module 305.

[0065] Ablation Experiment 1: Table 4 summarizes the performance of different module configurations, where GLFE represents the global-local feature extraction branch and SFE represents the structural feature extraction branch. When using GLFE alone, the model achieves an average Rank-1 accuracy of 91.4%, indicating that global and local appearance cues provide basic discriminative capabilities for gait recognition. Combining GLFE with the proposed structure-aware module 304 (SPM) improves the average accuracy to 93.4%, with a significant 3.5% improvement under the condition of wearing a coat (CL). This result demonstrates that SPM enhances the model's robustness to structural perturbations caused by drastic changes in clothing. Furthermore, introducing the multi-scale structural feature fusion module 305 (MSFF) further improves the accuracy to 93.9%, indicating that multi-scale fusion across adjacent angular positions helps obtain a more stable and discriminative structural representation.

[0066] Table 4 Performance Comparison of Different GaitSF Module Configurations

[0067] Ablation Experiment 2: Table 5 further reports the results of deploying the SFE branch to five representative gait recognition frameworks (including GaitSet, GaitPart, GaitBase, and GMSN). Under CL conditions, the integrated SFE improves performance by an average of 4.0% across all baselines, demonstrating that the proposed structural feature extractor exhibits good generalization and transferability across different gait recognition architectures.

[0068] Table 5 shows the results of deploying the SFE branch to five representative gait recognition frameworks.

[0069] Ablation Experiment 3: In the SPM module, the number of rays $R$ and the number of sampling points per ray $N$ are key factors affecting the ability to represent structural features. When $R$ is too large, the sampling intervals of adjacent angles will overlap significantly, which may weaken angle discrimination and reduce structural recognition ability. Conversely, when $R$ is too small, angle sampling is insufficient to fully capture changes in human geometric contours, thus limiting the ability to represent structures. Similarly, a value of $N$ that is too small may lead to unstable structural modeling, while a value of $N$ that is too large will introduce redundant features and noise, adversely affecting network optimization.

[0070] As shown in Table 6, setting the number of rays to R=64 and the number of sampling points per ray to N=5 achieves a good balance between preserving geometric information and suppressing feature redundancy, resulting in the highest recognition accuracy in our experiments. These results demonstrate that appropriate structural feature sampling in SPM helps improve recognition performance under complex appearance conditions.

[0071] Table 6. The Influence of Different Numbers of Rays and Number of Sampling Points per Ray in the SPM Module

[0072] Ablation Experiment 4: Table 7 shows the performance of MSFF under different kernel size configurations. When using kernels of size 3 or 5 alone, the average Rank-1 accuracy reaches 93.5% and 93.3%, respectively. This indicates that adaptively fusing angular structural features within a relatively small angular neighborhood can effectively integrate complementary local shape information and generate a more stable structural representation. In contrast, increasing the kernel size to 7 leads to a performance drop to 92.6%, suggesting that adaptive feature fusing over an excessively large angular range may introduce an oversmoothing effect, thereby weakening fine-grained angular discrimination and reducing discriminative ability.

[0073] Furthermore, employing a multi-scale configuration combining kernel sizes of 3 and 5 improves the average accuracy to 93.9%. This result demonstrates that adaptive fusion of multi-scale angular features enables the model to simultaneously utilize fine-grained local variations and a broader range of angular patterns, thereby generating a more comprehensive and discriminative structural representation. In contrast, when the multi-scale combination includes a kernel of size 7, performance drops to 92.8% or lower, indicating that an excessively large adaptive fusion range may contain redundant or noisy angular information, thus degrading representation quality.

[0074] Based on these observations, the combination of kernel sizes 3 and 5 achieves a good balance between effective adaptive angular feature fusion and preservation of fine-grained structural details, making it the most efficient and stable configuration for the MSFF module.

[0075] Table 7 Performance of MSFF under different kernel size configurations

[0076] Furthermore, this embodiment also included an occlusion experiment. In practical applications, pedestrian video sequences are frequently affected by occlusion, which can adversely impact gait recognition performance. For example... Figure 4 As shown, we simulated the impact of different occlusion regions on gait sequences. To evaluate the robustness of the proposed method under these conditions, we performed occlusion evaluation. Specifically, no occlusion samples were used during training; occlusion was introduced only during the testing phase by randomly placing square regions within each 64×44 pixel image frame.

[0077] like Figure 5 As shown, Figure 5 The results show how recognition accuracy changes with the size of the occlusion area. Accuracy gradually decreases under all three walking conditions as the occlusion area increases. Performance remains relatively stable when the occlusion size does not exceed 16 × 16 pixels. Under the NM condition, accuracy decreases by only 1.7%, still reaching 96.8%. Even when the occlusion area expands to 32 × 32 pixels, covering a large portion of the human body contour, the proposed method still maintains good discriminative ability. These observations demonstrate that the proposed method is robust to partial occlusion, thanks to its structural feature modeling strategy.

[0078] Example 4 like Figure 6As shown, this embodiment of the invention also provides an electronic device, the device including: a memory 606, a processor 605, and a computer program stored in the memory 606 and executable on the processor 605. The computer program is configured to implement the steps of the above-described gait recognition method and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0079] For details, see Figure 4 The present invention also provides an electronic device, including a bus 601, a transceiver 602, an antenna 603, a bus interface 604, a processor 605, and a memory 606.

[0080] The transceiver 602 is used to acquire unstructured data, which includes at least one of data obtained based on user input information and data obtained based on configuration file scanning. The processor 605 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 605 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. Processor 605 performs the various methods and processes described above, such as gait recognition methods.

[0081] exist Figure 4 In this document, a bus architecture (represented by bus 601) is used. Bus 601 may include any number of interconnected buses and bridges, linking various circuits including one or more processors represented by processor 605 and memory represented by memory 606. Bus 601 may also link various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. Bus interface 604 provides an interface between bus 601 and transceiver 602. Transceiver 602 may be a single element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by processor 605 is transmitted over a wireless medium via antenna 603, which further receives data and transmits data to processor 605.

[0082] Processor 605 manages bus 601 and general processing, and also provides various functions, including timing, peripheral interface, voltage regulation, power management, and other control functions. Memory 606 can be used to store data used by processor 605 during operation.

[0083] Optionally, the processor 605 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or a complex programmable logic device (CPLD).

[0084] Example 5 This invention also provides a storage medium, which is a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the steps of the gait recognition method described above and achieves the same technical effect. To avoid repetition, it will not be described again here. The computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0085] Example 6 This invention also provides a computer program product, which includes a computer program. When the computer program is executed by a processor, it implements the steps of the above-described gait recognition method and achieves the same technical effect. To avoid repetition, it will not be described again here.

[0086] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of the present invention is not limited to performing functions in the order discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0087] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0088] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of the present invention.

Claims

1. A gait recognition method, characterized in that, include: Define R different sampling angles, determine the center point of each sampling angle, where R is an integer greater than 1; at each sampling angle, take the corresponding center point as the origin of polar coordinates and sample radially to obtain human body structure data, and extract the structural features of the sampling angle based on the human body structure data; Gait recognition is performed based on the structural features at the R different sampling angles.

2. The gait recognition method according to claim 1, characterized in that, At each of the aforementioned sampling angles, sampling is performed radially with the corresponding center point as the origin of polar coordinates to obtain human body structure data, including: At each sampling angle, N sampling points are set radially with the corresponding center point as the origin of polar coordinates, where N is an integer greater than 1; The center point and the N sampling points are sampled respectively to obtain N+1 structural point data, and the N+1 structural point data constitute the human body structure data; Feature extraction is performed on each of the N+1 structural point data to obtain N+1 structural point features; The structural features are generated by weighting and linearly combining the N+1 structural point features according to their respective weights.

3. The gait recognition method according to claim 1, characterized in that, The gait recognition based on the structural features at the R different sampling angles includes: The structural features from the R different sampling angles are fused to obtain the final structural features; Gait recognition is performed based on the final structural features.

4. The gait recognition method according to claim 3, characterized in that, Also includes: Gait features are extracted, including global features representing the overall contour movement pattern of the human body and local features representing the movement changes of body parts; In the gait recognition based on the final structural features, gait recognition is performed according to the gait features and the final structural features.

5. The gait recognition method according to claim 4, characterized in that, The gait recognition based on the gait features and the final structural features includes: The gait features are fused with the final structural features to obtain fused features; The fused features are sequentially processed by temporal pooling, horizontal pyramid pooling, a first fully connected layer and batch normalization, and a second fully connected layer to output the gait identity category.

6. The gait recognition method according to claim 4, characterized in that, The extraction of gait features includes global features representing the overall contour movement pattern of the human body and local features representing changes in the movement of body parts, including: A sequence of gait images to be identified is obtained, and features are extracted from the gait image sequence using a three-dimensional convolutional network to obtain spatiotemporal features. The gait features are extracted based on the spatiotemporal features.

7. A gait recognition system, characterized in that, include: The structure perception module is used to define R different sampling angles, determine the center point of each sampling angle, where R is an integer greater than 1; at each sampling angle, sampling is performed radially with the corresponding center point as the origin of polar coordinates to obtain human body structure data, and the structural features of the sampling angle are extracted based on the human body structure data. The recognition module is used to perform gait recognition based on the structural features at the R different sampling angles.

8. An electronic device, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the gait recognition method as described in any one of claims 1 to 6.

9. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the gait recognition method as described in any one of claims 1 to 6.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the steps of the gait recognition method as described in any one of claims 1 to 6.

Images