Variable-length time series gait recognition system and method based on regional perception cascaded convolution

Abstract

This invention discloses a variable-length temporal gait recognition system and method based on region-aware cascaded convolutions, belonging to the fields of computer vision and biometric recognition technology. The invention divides gait features into two ends and a middle region along the height direction using region partitioning units. For the two ends, a deep temporal extraction branch is configured, employing three levels of cascaded convolutions to capture long-period complex dynamics. For the middle region, a shallow temporal extraction branch is configured, employing two levels of cascaded convolutions to extract stationary features, forming an asymmetric architecture that is deep at the ends and shallow in the middle. Each cascaded convolution unit integrates a temporal attention enhancement circuit, achieving weighted enhancement of keyframes and suppression of noisy frames through parallel attention generation and feature preservation paths. Through the synergistic effect of the region-aware asymmetric design and the attention enhancement mechanism, the system accurately adapts to the differences in movement frequency of different parts of the human body, significantly improving the accuracy and robustness of gait recognition while ensuring computational efficiency.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and biometric recognition technology, and in particular relates to a variable-length temporal gait recognition system and method based on region-aware cascaded convolution that can perform asymmetric temporal modeling of the differences in movement frequency in different anatomical regions of the human body. Background Technology

[0002] Gait recognition is a biometric identification technology that identifies individuals by analyzing their walking posture. Compared to contact-based or short-range identification methods such as facial recognition and fingerprint recognition, gait recognition has significant advantages such as long-range, non-contact, difficult to spoof, and able to work even at low resolution. Therefore, it shows broad application prospects in fields such as intelligent security, video surveillance, criminal investigation, and access control.

[0003] Currently, mainstream gait recognition methods typically employ 3D convolutional neural networks (3DCNNs) to extract spatiotemporal features from input gait contour sequences. Although existing deep learning methods have achieved high recognition accuracy on publicly available datasets, they still face several significant technical bottlenecks in practical, complex scenarios: First, existing methods neglect the physical differences in motion frequencies across different parts of the human body. Current techniques generally employ a "one-size-fits-all" global temporal modeling strategy, processing all regions of a human image (from head to toe) with the same number of layers and the same size of convolutional kernels. However, according to biomechanical principles, the motion patterns of different parts of the body during walking are distinctly different: the feet and head typically involve high-frequency swaying or complex periodic changes, requiring deeper networks to capture long-term temporal dependencies; while the relative motion of the torso and thighs is relatively smooth and slow. Existing methods fail to model these regional differences specifically, resulting in insufficient feature capture in high-frequency regions or wasted computational resources and overfitting in low-frequency regions.

[0004] Second, the single and fixed temporal receptive field limits the expression of long-term features. To control computational complexity, most existing methods use small and fixed-size convolutional kernels (such as using only k=3 kernels) for local temporal aggregation. This limited receptive field makes it difficult to cover the complete gait cycle, causing the model to only capture instantaneous micro-movements and lose long-term macro-discriminative information such as stride rhythm and center of gravity fluctuations. Especially when dealing with edge regions with complex periodicity (such as limb extremities), short-term features often lack sufficient robustness.

[0005] Third, there is a lack of effective frame-level noise suppression mechanisms. In actual surveillance videos, gait sequences often contain occlusions, shadow interference, or incomplete invalid frames. Traditional temporal aggregation methods (such as simple max pooling or average pooling) process the information of all frames equally or retain only the strongest response, failing to intelligently filter out the most recognizable key frames. This results in the final generated feature vector being easily contaminated by noisy frames, reducing the accuracy of recognition.

[0006] Furthermore, in existing technologies, the processing modules are often independent of each other and lack collaborative design for human motion characteristics. This results in a disconnect in information transmission between feature extraction and subsequent processing, preventing the formation of an organic whole and limiting further improvement in recognition performance.

[0007] In summary, there is an urgent need for a gait recognition system and method that can perform differentiated modeling based on the characteristics of human anatomical regions, ensure computational efficiency while taking into account the extraction of long and short temporal features, and has keyframe enhancement capabilities. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the prior art and provide a gait recognition system and method that can accurately adapt to the differences in movement frequency of different parts of the human body, achieve collaborative extraction of long and short temporal features through asymmetric cascaded convolution, and integrate a frame-level attention enhancement mechanism.

[0009] To achieve the above objectives, the present invention provides the following technical solution: A variable-length temporal gait recognition method based on region-aware concatenated convolution includes the following steps: S1: Obtain the gait contour sequence of the pedestrian to be identified, and perform preprocessing to obtain a three-dimensional feature map; S2: Extract spatial features from the three-dimensional feature map to obtain a spatial feature tensor; S3: Input the spatial feature tensor into the differential temporal aggregation module for temporal feature extraction, including: S31: The spatial feature tensor is divided into multiple semantic region channels in the height direction by the region division unit. The semantic region channels include at least the two end region channels corresponding to the human head and feet, and the middle region channel corresponding to the human torso and thighs. S32: For the two end region channels, temporal features are extracted through a deep temporal extraction branch. The deep temporal extraction branch adopts a three-level cascaded convolutional structure, which sequentially connects the first convolutional unit, the second convolutional unit, and the third convolutional unit. The third convolutional unit is configured with a large-size convolutional kernel to cover the long temporal receptive field. S33: For the intermediate region channel, temporal features are extracted through a shallow temporal extraction branch. The shallow temporal extraction branch adopts a two-level cascaded convolutional structure, which only contains the first and second convolutional units connected in series, and does not contain the third convolutional unit. S34: The features output by the deep temporal extraction branch and the features output by the shallow temporal extraction branch are fused together by the fusion aggregator to obtain aggregated temporal features; S4: Input the aggregated temporal features into the feature mapping module, perform spatial pooling and dimension transformation to obtain a fixed-length identity feature vector; S5: Output the fixed-length identity feature vector through the identity feature output terminal for identity recognition.

[0010] Furthermore, the kernel size of the first convolutional unit is configured as k=3, the kernel size of the second convolutional unit is configured as k=5, and the kernel size of the third convolutional unit is configured as k=7. The deep temporal extraction branch forms a deep temporal modeling pathway with progressively expanding receptive fields through a three-level cascade of k=3, 5, and 7, in order to capture the complex dynamic features of head micro-movements and foot strides. This incremental convolutional kernel design enables the shallow layer to capture instantaneous movements, the middle layer to capture local swaying, and the deep layer to cover the complete gait cycle, forming an organic fusion of multi-scale temporal features.

[0011] Furthermore, the first, second, and third convolutional units each integrate a temporal attention enhancement circuit, which includes a parallel temporal attention generation path and a feature preservation path. The temporal attention generation path generates weight coefficients reflecting frame-level importance based on the input temporal features; the feature preservation path preserves the texture information of the input temporal features; the temporal attention enhancement circuit also includes a multiplication node, which performs a dot product operation between the weight coefficients and the features output by the feature preservation path to output the enhanced temporal features.

[0012] Furthermore, the temporal attention generation path is specifically implemented as follows: the input temporal features are channel-compressed using a one-dimensional convolutional layer, and then an attention mask between 0 and 1 is generated using a sigmoid activation function, which serves as the frame-level importance weight coefficient. The feature preservation path is specifically implemented as follows: the input temporal features are pooled using parallel average pooling and max pooling layers, and the pooling results are summed to preserve complete texture features. This parallel path design ensures that attention generation and feature preservation do not interfere with each other, and at the hardware level, gate suppression of invalid frame signals and gain amplification of key frame signals are achieved.

[0013] Furthermore, the region segmentation unit divides the spatial feature tensor into four semantic region channels along the height direction, corresponding to the head region, torso region, thigh region, and foot region of the human body, respectively; among them, the head region and foot region are divided into two end region channels, and the torso region and thigh region are divided into two middle region channels. This segmentation method is based on human anatomical structure, ensuring that the physical meaning of subsequent differential processing is clear.

[0014] Furthermore, the fusion operation performed by the fusion aggregator is tensor splicing, which splices the features output by the deep temporal extraction branch with the features output by the shallow temporal extraction branch in the height dimension to restore a complete full-body feature map, ensuring the integrity and spatial consistency of the features.

[0015] The present invention also provides a variable-length temporal gait recognition system based on region-aware concatenated convolution, comprising: Video capture equipment is used to acquire gait contour sequences of pedestrians to be identified; Memory, used to store computer programs; A processor that implements the above-described gait recognition method when executing the computer program; The output interface is used to output the identity verification results.

[0016] Furthermore, the processor includes: An input interface, connected to the video acquisition device, is used to receive gait contour sequences; The spatial feature extraction module is used to extract spatial features from gait contour sequences and output a spatial feature tensor. The differentiated temporal aggregation module is used to perform region-aware asymmetric temporal feature extraction on spatial feature tensors, including: a region partitioning unit, which divides the spatial feature tensor into two end region channels and a middle region channel in the height direction; The deep temporal extraction branch is connected to the two-end region channels and adopts a three-level cascaded convolutional structure, which sequentially connects the first convolutional unit, the second convolutional unit and the third convolutional unit; The shallow temporal extraction branch is connected to the intermediate region channel and adopts a two-level cascaded convolutional structure, which contains only the first and second convolutional units connected in series. A fusion aggregator is used to fuse the output features of the deep temporal extraction branch and the shallow temporal extraction branch; a feature mapping module is connected to the differential temporal aggregation module and is used to map the aggregated temporal features into a fixed-length identity feature vector. The identity feature output terminal is connected to the feature mapping module and is used to output a fixed-length identity feature vector.

[0017] Furthermore, the first convolutional unit, the second convolutional unit, and the third convolutional unit all integrate a temporal attention enhancement circuit, which includes: The temporal attention generation path consists of a one-dimensional convolutional layer and a sigmoid activation function, used to generate frame-level attention weights; The feature preservation path consists of parallel average pooling layers and max pooling layers, used to preserve texture features; The multiplication node, connected to the output of the temporal attention generation path and the feature preservation path, is used to perform dot product between the attention weights and the texture features, and output the enhanced temporal features.

[0018] Compared with the prior art, the present invention has the following beneficial effects: 1. High regional perception accuracy: Through an asymmetrical architecture design with deep ends and a shallow middle, it accurately adapts to the movement patterns of different parts of the human body. The deep branches capture high-frequency details of head micro-movements and foot strides, while the shallow branches stabilize the overall posture of the torso and upper arms. This effectively solves the feature confusion problem caused by the one-size-fits-all approach of traditional methods, and achieves accurate modeling of the physical essence of human movement.

[0019] 2. Strong multi-scale temporal perception capability: The deep branch adopts an incremental cascaded convolution design with k=3, 5, and 7, and the receptive field is gradually expanded, enabling the model to simultaneously perceive temporal information at three scales: instantaneous action, local sway, and complete gait cycle. This breaks through the limitations of the traditional short time window and significantly improves the expressive ability of long-cycle gait prosodic features.

[0020] 3. Synergistic effect of keyframe enhancement and noise suppression: The temporal attention enhancement circuit integrated within each convolutional unit achieves adaptive weighted enhancement of keyframes and gating suppression of noisy frames through parallel attention generation and feature preservation paths. This mechanism ensures that the final generated identity feature vector is mainly contributed by high-discrimination frames, effectively improving recognition robustness.

[0021] 4. Significantly optimized computational efficiency: Deep convolution calculations are omitted in the torso and thigh regions with low information content, and computational resources are concentrated in the high-frequency regions with rich information content. While improving recognition performance, the ineffective consumption of computing power is avoided, and an optimized balance between accuracy and efficiency is achieved.

[0022] 5. Full-process collaborative linkage: From region partitioning and asymmetric concatenated convolution to temporal attention enhancement, each step of this invention has an inherent linkage relationship: the result of region partitioning directly determines the selection of branches, the depth design of branches adapts to the motion frequency of the regions, and the attention mechanism works collaboratively within each concatenated unit. This full-chain collaborative design makes each module no longer an isolated step, but a tightly coupled organic whole, producing a collaborative technical effect of 1+1>2. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the overall architecture of the gait recognition system provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the internal logic structure of the differentiated timing aggregation module of the present invention; Figure 3 yes Figure 2 A schematic diagram of the timing attention enhancement circuit inside a cascaded convolutional unit. Detailed Implementation

[0024] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0025] Example 1: System Hardware Structure This embodiment provides a variable-length temporal gait recognition system based on region-aware concatenated convolution, the overall architecture of which is as follows: Figure 1 As shown, the system mainly includes video acquisition equipment, memory, processor, and output interfaces. The video acquisition equipment (such as a high-definition surveillance camera or infrared camera) captures continuous contour sequences of pedestrians during their walking process and transmits the raw video data to the processor. The processor, connected to the memory, loads and executes the computer program stored therein to implement the complete gait recognition process. The output interface is used to send the recognition results to an access control system, monitoring platform, or remote server. In practical deployments, this system can run on embedded devices, edge computing terminals, or cloud servers, demonstrating good engineering applicability.

[0026] The processor integrates multiple sequentially connected functional modules, forming an end-to-end feature processing path: Input interface: Physically connected to the video acquisition device for receiving gait contour sequences in real time.

[0027] Spatial feature extraction module: Performs preprocessing on the received gait contour sequence, such as background removal, size normalization and frame alignment, and extracts spatial features through a three-dimensional convolutional neural network backbone network (such as 3DCNN), outputting a spatial feature tensor, with dimensions represented as C×T×H×W (number of channels × number of time frames × feature map height × width).

[0028] Differentiated time-series aggregation module: The core module of this invention, its internal logic structure is as follows Figure 2 As shown, this module is used for region-aware asymmetric temporal feature extraction from spatial feature tensors. The module includes: Region partitioning unit: The input spatial feature tensor is divided into four semantic region channels along the height dimension, corresponding to the head region (above H / 4), trunk region (H / 4 to H / 2), thigh region (H / 2 to 3H / 4), and foot region (below 3H / 4). The head and foot regions are divided into end-point channels, while the trunk and thigh regions are divided into middle-point channels. This partitioning method is based on human anatomy, ensuring the clear physical meaning of subsequent differential processing.

[0029] The deep temporal extraction branch connects the two end regions (head and feet) and employs a three-level cascaded convolutional structure, sequentially connecting the first, second, and third convolutional units. The first convolutional unit uses a kernel with k=3 to capture instantaneous movements; the second convolutional unit uses a kernel with k=5 to capture local swaying; and the third convolutional unit uses a kernel with k=7 to cover the complete gait cycle. This incremental kernel design forms a deep temporal modeling pathway with progressively expanding receptive fields to capture the complex dynamic features of head micro-movements and foot strides.

[0030] Shallow temporal extraction branch: Connected to the middle region channel (torso and thigh), it adopts a two-level cascaded convolutional structure, containing only the first convolutional unit (k=3) and the second convolutional unit (k=5) connected in series, without the third convolutional unit. This design utilizes a shallow number of network layers and a moderate temporal receptive field to efficiently extract relatively smooth motion features of the torso and thigh regions, avoiding the introduction of redundant noise.

[0031] Fusion Aggregator: The features output by the deep temporal extraction branch and the features output by the shallow temporal extraction branch are tensor-concatenated in the height dimension to restore the complete full-body feature map and output the aggregated temporal features.

[0032] Feature Mapping Module: Connected to the Differentiated Temporal Aggregation Module, this module maps aggregated temporal features into fixed-length identity feature vectors. It typically includes horizontal pyramid pooling units and fully connected mapping units, compressing variable-length temporal features into fixed-dimensional identity feature vectors through spatial pooling and dimensionality transformation.

[0033] Identity feature output end: connected to the feature mapping module, used to output a fixed-length identity feature vector for downstream database comparison system to complete identity recognition.

[0034] Example 2: Internal Structure of Cascaded Convolutional Units To further improve the purity of feature extraction, this embodiment refines the internal structure of the first, second, and third convolutional units, such as... Figure 3 As shown. Each cascaded convolutional unit integrates a time-series attention enhancement circuit, which includes two parallel signal paths: The temporal attention generation path consists of a cascaded one-dimensional convolutional layer and a sigmoid activation function. The one-dimensional convolutional layer performs channel compression on the input temporal features, reducing multi-channel features to a single-channel frame-level response. The sigmoid activation function maps the compressed response to a value between 0 and 1, generating an attention mask representing frame-level importance. This mask reflects the contribution of each frame to identity recognition; noisy frames are assigned weights close to 0, and key frames are assigned weights close to 1.

[0035] Feature Preservation Path: Consists of parallel average pooling layers and max pooling layers. The average pooling layer preserves the global statistical information of the features, while the max pooling layer preserves the local peak information of the features. Adding the outputs of the two layers can completely preserve the texture information of the input features, avoiding information loss that may be caused by the attention mechanism.

[0036] The outputs of the two paths are connected to a multiplication node, which performs frame-by-frame dot product operations on the attention mask and texture features to achieve gating suppression of invalid frames and feature gain of key frames, and finally outputs the enhanced temporal features.

[0037] This parallel path design ensures that attention generation and feature preservation do not interfere with each other, and at the hardware level, it achieves adaptive suppression of invalid frame signals and gain amplification of key frame signals, significantly improving the quality of features.

[0038] Example 3: Gait Recognition Method Flowchart This embodiment provides a variable-length temporal gait recognition method based on region-aware concatenated convolution, specifically including the following steps: S1: Obtain the gait contour sequence of the pedestrian to be identified, and perform preprocessing such as background removal, size normalization, and frame alignment to obtain the preprocessed gait sequence; S2: Spatial features are extracted through a 3D convolutional neural network backbone network, and a spatial feature tensor with dimensions C×T×H×W is output. S3: Input the spatial feature tensor into the differential temporal aggregation module to extract temporal features; S31: The spatial feature tensor is divided into four semantic region channels in the height direction by the region segmentation unit: head region, torso region, thigh region, and foot region; among which the head and feet are the two end region channels, and the torso and thigh are the middle region channels. S32: For the two end regions, temporal features are extracted through a deep temporal extraction branch: sequentially through the first convolutional unit (k=3, with temporal attention enhancement), the second convolutional unit (k=5, with temporal attention enhancement), and the third convolutional unit (k=7, with temporal attention enhancement), and output the enhanced features of the two end regions; S33: For the middle region channel, temporal features are extracted through a shallow temporal extraction branch: the middle region enhanced features are output by sequentially passing through the first convolutional unit (k=3, with temporal attention enhancement) and the second convolutional unit (k=5, with temporal attention enhancement); S34: By using a fusion aggregator, the enhanced features of the two ends and the enhanced features of the middle region are tensor-joined in the height dimension to restore the complete full-body aggregated temporal features; S4: Input the aggregated temporal features into the feature mapping module, and then pass them through horizontal pyramid pooling and fully connected mapping to obtain a fixed-length identity feature vector; S5: Output a fixed-length identity feature vector through the identity feature output terminal, compare it with the registered features in the database, and complete the identity recognition.

[0039] Example 4: Detailed Explanation of Collaborative Relationships To more clearly demonstrate the collaborative design of this invention, the linkage relationships between the modules are described in detail below: The linkage between region segmentation and branch selection: Based on human anatomical structure, the region segmentation unit precisely divides the feature map into four semantic regions in the height dimension. This segmentation result directly determines the selection of subsequent branches: the head and foot regions are routed to the deep temporal extraction branch, while the torso and thigh regions are routed to the shallow temporal extraction branch. This linkage design of segmentation and routing ensures that the physical meaning of differentiated processing is clear and avoids errors that may be caused by manually specifying regions.

[0040] Linkage between branch depth and motion frequency: The deep temporal extraction branch employs a three-level cascade (k=3, 5, 7), with the receptive field expanding progressively to precisely match the high-frequency, complex motion characteristics of the head and feet; the shallow temporal extraction branch employs a two-level cascade (k=3, 5), with a moderate receptive field to precisely match the low-frequency, smooth motion characteristics of the torso and thighs. This linkage design between motion frequency and network depth allows computational resources to be precisely allocated to information-rich areas, achieving an optimized balance between accuracy and efficiency.

[0041] The cascaded convolution and attention enhancement mechanism works in tandem: each cascaded convolutional unit integrates a temporal attention enhancement circuit, which operates collaboratively within each unit. The attention mechanism of shallow units (k=3) primarily suppresses transient noise frames; the attention mechanism of mid-level units (k=5) focuses on local sway keyframes; and the attention mechanism of deep units (k=7) enhances representative frames within the complete gait cycle. This cascaded depth-attention granularity linkage design enables multi-level keyframe filtering from micro to macro perspectives, significantly improving feature quality.

[0042] Linkage between fusion aggregation and subsequent mapping: The fusion aggregator uses tensor splicing operation to restore the features of the two ends and the middle region into a complete full-body feature map in the height dimension. This ensures that the subsequent feature mapping module can perform pooling operation based on spatially complete features, avoiding the loss of spatial information due to feature dispersion.

[0043] Through the above-mentioned linkage design, the modules of this invention are no longer isolated functional units, but form a tightly coupled organic whole, which together improves the performance and robustness of gait recognition.

[0044] Example 5: Experimental Verification and Effect Comparison To further verify the technical effectiveness of this invention, a comparative experiment was conducted on the publicly available gait dataset CASIA-B. CASIA-B contains 124 pedestrians and three walking conditions: normal (NM), backpack (BG), and wearing a coat (CL). The experimental setup is as follows: using the same backbone network (3DCNN), the following methods were compared: Method A (baseline): Globally uniform three-level concatenated convolution (k=3, 5, 7), no region partitioning, no attention enhancement; Method B: It involves region division (deep at both ends, light in the middle), but does not enhance attention. Method C: It has attention enhancement, but no region division (it uses a three-level cascade globally). Method D (this invention): Region partitioning + asymmetric cascading + attention enhancement.

[0045] The experimental results are shown in Table 1: Table 1 Comparison of recognition accuracy of different methods on the CASIA-B dataset. As shown in Table 1, the method of this invention achieves an average recognition accuracy of 91.3% in three scenarios: normal, backpack, and wearing a coat, significantly outperforming the baseline method's 85.9%, with an average improvement of 5.4 percentage points. In the most challenging scenario of wearing a coat (CL), the accuracy of this invention reaches 81.1%, a significant improvement of 8.0 percentage points compared to the baseline method (73.1%); in the backpack scenario (BG), the accuracy reaches 94.6%, an improvement of 5.3 percentage points compared to the baseline method (89.3%); and in the normal scenario (NM), the accuracy reaches 98.3%, also significantly outperforming other comparative methods. This significant improvement is attributed to the collaborative linkage design of this invention: the region division and asymmetric cascading linkage accurately adapt to the motion characteristics of different parts, and the attention enhancement mechanism effectively filters keyframes through collaborative work within each cascading unit. The organic combination of these two elements produces a synergistic effect of 1+1>2, rather than a simple additive effect.

[0046] Therefore, this invention has outstanding substantive features and significant progress compared to the prior art, mainly reflected in the following aspects: Non-obviousness: While existing technologies include concepts such as region-based processing, multi-scale convolution, and attention mechanisms, combining these modules in the specific manner described in this invention and establishing precise linkages (e.g., region partitioning directly determines branch selection, branch depth precisely matches motion frequency, and attention mechanisms work collaboratively within cascaded units) is not a conventional choice for those skilled in the art. In particular, the asymmetric cascaded architecture with "deep at both ends and shallow in the middle," and the long-period modeling using large k=7 convolutional kernels for the head and foot regions, require a deep understanding of the technical problems addressed in this invention and the biomechanical characteristics of human movement to propose.

[0047] Collaborative Solution to Technical Challenges: This invention simultaneously addresses three major challenges in gait recognition—ignoring regional motion frequency differences, a single and fixed temporal receptive field, and the lack of frame-level noise suppression mechanisms. More importantly, the interrelationship between modules ensures that these solutions are not isolated but mutually reinforcing: the linkage between region segmentation and branch depth enables precise allocation of computational resources, and the linkage between cascaded convolution and attention enhancement enables multi-level keyframe selection, ultimately resulting in a significant improvement in overall performance.

[0048] Significant technical effect: Experimental data shows that the recognition accuracy of this invention is significantly better than existing methods in complex scenarios (such as clothing changes and occlusion), with an average improvement of 6.7 percentage points. This improvement cannot be achieved by simple parameter adjustments or conventional improvements, but rather stems from the innovative design of the overall technical solution and the synergistic interaction between modules.

[0049] Specialized design for specific technical problems: This invention is not a general video recognition method, but is specifically designed to address practical issues in gait recognition, such as the movement characteristics of human body parts, gait cycle length, and frame-level noise interference. For example, region segmentation is based on human anatomical structure, the large convolutional kernel with k=7 is designed to cover the complete gait cycle, and the attention enhancement circuit is used to suppress occlusion and shadow interference. This targeted design makes this invention stand out in the specific field of gait recognition.

[0050] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A variable-length temporal gait recognition method based on region-aware concatenated convolution, characterized in that, Includes the following steps: S1: Obtain the gait contour sequence of the pedestrian to be identified, and perform preprocessing to obtain a three-dimensional feature map; S2: Extract spatial features from the three-dimensional feature map to obtain a spatial feature tensor; S3: Input the spatial feature tensor into the differential temporal aggregation module for temporal feature extraction, including: S31: The spatial feature tensor is divided into multiple semantic region channels in the height direction by the region division unit. The semantic region channels include at least the two end region channels corresponding to the human head and feet, and the middle region channel corresponding to the human torso and thighs. S32: For the two end region channels, temporal features are extracted through a deep temporal extraction branch. The deep temporal extraction branch adopts a three-level cascaded convolutional structure, which sequentially connects the first convolutional unit, the second convolutional unit, and the third convolutional unit. The third convolutional unit is configured with a large-size convolutional kernel to cover the long temporal receptive field. S33: For the intermediate region channel, temporal features are extracted through a shallow temporal extraction branch. The shallow temporal extraction branch adopts a two-level cascaded convolutional structure, which only contains the first and second convolutional units connected in series, and does not contain the third convolutional unit. S34: The features output by the deep temporal extraction branch and the features output by the shallow temporal extraction branch are fused together by the fusion aggregator to obtain aggregated temporal features; S4: Input the aggregated temporal features into the feature mapping module, perform spatial pooling and dimension transformation to obtain a fixed-length identity feature vector; S5: Output the fixed-length identity feature vector through the identity feature output terminal for identity recognition.

2. The gait recognition method according to claim 1, characterized in that, The kernel size of the first convolutional unit is configured as k=3, the kernel size of the second convolutional unit is configured as k=5, and the kernel size of the third convolutional unit is configured as k=7. The deep temporal extraction branch forms a deep temporal modeling pathway with progressively expanding receptive fields through a three-level cascade of k=3, 5, and 7, in order to capture the complex dynamic features of head micro-movements and foot strides.

3. The gait recognition method according to claim 1, characterized in that, The first, second, and third convolutional units each integrate a temporal attention enhancement circuit, which includes a parallel temporal attention generation path and a feature preservation path. The temporal attention generation path is used to generate weight coefficients that reflect frame-level importance based on the input temporal features; the feature preservation path is used to preserve the texture information of the input temporal features. The temporal attention enhancement circuit also includes a multiplication node, which is used to perform a dot product operation on the weight coefficients and the features output by the feature preservation path to output the enhanced temporal features.

4. The gait recognition method according to claim 3, characterized in that, The specific implementation of the temporal attention generation path is as follows: channel compression of the input temporal features is performed through a one-dimensional convolutional layer, and then an attention mask between 0 and 1 is generated by the Sigmoid activation function, which serves as the frame-level importance weight coefficient. The specific implementation of the feature preservation path is as follows: the input temporal features are pooled by parallel average pooling layers and max pooling layers, and the pooling results are added together to preserve the complete texture features.

5. The gait recognition method according to claim 1, characterized in that, The region partitioning unit divides the spatial feature tensor into four semantic region channels in the height direction, corresponding to the head region, torso region, thigh region, and foot region of the human body, respectively; among them, the head region and foot region are divided into two end region channels, and the torso region and thigh region are divided into middle region channels.

6. The gait recognition method according to claim 1, characterized in that, The fusion aggregator performs a tensor splicing operation, which splices the features output by the deep temporal extraction branch with the features output by the shallow temporal extraction branch in the height dimension to restore a complete full-body feature map.

7. The gait recognition method according to claim 1, characterized in that, The preprocessing of the gait contour sequence includes background removal, size normalization, and frame alignment operations; the spatial feature extraction is achieved through a three-dimensional convolutional neural network backbone network.

8. A variable-length temporal gait recognition system based on region-aware concatenated convolution, characterized in that, include: Video capture equipment is used to acquire gait contour sequences of pedestrians to be identified; Memory, used to store computer programs; A processor that, when executing the computer program, implements the gait recognition method as described in any one of claims 1 to 7; The output interface is used to output the identity verification results.

9. The gait recognition system according to claim 8, characterized in that, The processor includes: An input interface, connected to the video acquisition device, is used to receive gait contour sequences; The spatial feature extraction module is used to extract spatial features from gait contour sequences and output a spatial feature tensor. The differentiated temporal aggregation module is used to perform region-aware asymmetric temporal feature extraction on spatial feature tensors, including: a region partitioning unit, which divides the spatial feature tensor into two end region channels and a middle region channel in the height direction; The deep temporal extraction branch is connected to the two-end region channels and adopts a three-level cascaded convolutional structure, which sequentially connects the first convolutional unit, the second convolutional unit and the third convolutional unit; The shallow temporal extraction branch is connected to the intermediate region channel and adopts a two-level cascaded convolutional structure, which contains only the first and second convolutional units connected in series. A fusion aggregator is used to fuse the output features of the deep temporal extraction branch and the shallow temporal extraction branch; a feature mapping module is connected to the differential temporal aggregation module and is used to map the aggregated temporal features into a fixed-length identity feature vector. The identity feature output terminal is connected to the feature mapping module and is used to output a fixed-length identity feature vector.

10. The gait recognition system according to claim 9, characterized in that, The first, second, and third convolutional units each integrate a temporal attention enhancement circuit, which includes: The temporal attention generation path consists of a one-dimensional convolutional layer and a sigmoid activation function, used to generate frame-level attention weights; The feature preservation path consists of parallel average pooling layers and max pooling layers, used to preserve texture features; The multiplication node, connected to the output of the temporal attention generation path and the feature preservation path, is used to perform dot product between the attention weights and the texture features, and output the enhanced temporal features.

Images