An event-based gait recognition method

Abstract

The application discloses an event-based gait recognition method, comprising the following steps: 1, acquiring an event stream and a synchronous RGB image, and constructing short-term and long-term event slices; 2, constructing a dynamic motion stream and a static shape stream, wherein the dynamic motion stream is composed of a pulse gating network and a parallel pulse expert, and dynamic motion features are extracted from the short-term slice; 3, the static shape stream can extract RGB image features by using a frozen DINOv2 teacher model, guide a static encoder to learn shape features from the long-term slice, and calculate a cross-modal alignment loss; 4, the dynamic and static features are fused into embedding features through a fusion module; 5, an identification module predicts identity according to the embedding features, and the network is trained in combination with cross-entropy, a triple loss and an alignment loss. The application combines the high dynamic characteristics of the event stream and the semantic information of the RGB image, and improves the robustness and accuracy of gait recognition in a complex scene.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and biometric recognition technology, specifically to an event-flow-based gait recognition method and model training method for robust scenarios. Background Technology

[0002] Gait recognition aims to identify individuals based on their walking patterns. It offers advantages such as long-range, non-contact, and difficulty in spoofing, making it promising for applications in security monitoring and medical rehabilitation. Traditional gait recognition methods primarily rely on video frames captured by RGB cameras. However, RGB cameras often struggle to extract reliable visual features (such as contours or skeletons) in uncontrolled environments like low light, strong lighting changes, or cluttered backgrounds, leading to a sharp decline in recognition performance. While LiDAR-based gait recognition methods exhibit strong robustness, their high cost and energy consumption limit large-scale deployment.

[0003] In contrast, event cameras (or dynamic vision sensors, DVS) are bio-inspired sensors that asynchronously record pixel-level changes in brightness (events), offering extremely high temporal resolution (microseconds) and dynamic range (>120dB). Event cameras naturally capture rich dynamic motion information and filter out redundant static backgrounds, making event-based gait recognition a potential solution for identity recognition under complex lighting conditions.

[0004] However, existing event-based gait recognition methods have not fully tapped the potential of event data and still have the following key limitations: (1) Waste of temporal resolution and motion blur: Existing methods usually simply aggregate the event stream within a time window into static "event frames" or "voxel grids". This approach will lose high-frequency fine motion dynamic information under long exposure windows, resulting in motion blur effect similar to that of traditional cameras, which destroys the key dynamic cues in gait recognition.

[0005] (2) Difficulty in feature extraction due to spatial sparsity: Event data is naturally sparse in space (triggered only at brightness changes). Existing convolutional neural networks (CNNs) have difficulty extracting dense and discriminative body shape features from such sparse inputs, resulting in insufficient modeling ability for pedestrian appearance.

[0006] (3) Single neuron dynamics cannot adapt to complex scenes: In dynamic stream processing, existing spiking neural networks (SNNs) typically use a fixed membrane time constant. However, in real-world scenes, lighting and motion speed vary (e.g., from bright to dark, or walking at fast or slow speeds). A single time constant cannot simultaneously adapt to event streams with high-frequency triggering (strong light / fast motion) and sparse triggering (weak light / slow motion), thus limiting the robustness of the model.

[0007] The aforementioned problems limit the effectiveness of event-based gait recognition in real-world complex scenarios. There is an urgent need to design a gait recognition framework that can balance high-frequency dynamic capture and dense shape recovery, and can adapt to different environmental conditions. Summary of the Invention

[0008] To overcome the shortcomings of existing event-based gait recognition methods, such as loss of dynamic information, sparse spatial features, and poor adaptability to changing environments, this invention provides an event-based gait recognition method. It aims to model fine motion and dense shape separately through a dual-stream architecture and utilize a hybrid expert mechanism and cross-modal alignment technology, thereby significantly improving the gait recognition accuracy in complex lighting scenarios.

[0009] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The gait recognition method based on events of this invention is characterized by the following steps: Step 1: Use the event camera to collect standard event streams and positive sample event streams for the same real identity, as well as negative sample event streams for different real identities; The standard event stream, positive sample event stream, and negative sample event stream are respectively divided into... One exposure window; The standard events, positive sample events, and negative sample events under the m-th exposure window are respectively divided into... The standard events, positive sample events, and negative sample events in each interval are aggregated into short-term standard event slices, short-term positive sample event slices, and short-term negative sample event slices, respectively. Let the short-term standard event slice corresponding to the k-th interval under the m-th exposure window be denoted as . Let the slice of short-term positive sample events corresponding to the k-th interval under the m-th exposure window be denoted as Let the slice of short-term negative sample events corresponding to the k-th interval under the m-th exposure window be denoted as ; Aggregate the standard events under the m-th exposure window into the m-th long-term standard event slice. The positive sample events under the m-th exposure window are aggregated into the m-th long-term positive sample event slice. The negative sample events under the m-th exposure window are aggregated into the m-th long-term negative sample event slice. ; Obtain the RGB image captured by a regular camera corresponding to the m-th exposure window of the standard event stream and perform grayscale processing to obtain the intensity image under the m-th exposure window. ; One-hot encoding of the real identity corresponding to the standard event stream =[ ,..., ,… ], For the r-th real identity code, Indicates the total number of identities; Step 2: Construct an EventGait dual-stream network, including: dynamic motion flow and static shape flow; Step 2.1, the dynamic motion flow consists of a lightweight pulse-gated network and A team of parallel pulse experts was formed, and [the team] conducted [the following]... The process is performed to obtain the k-th standard dynamic motion feature under the m-th exposure window. ; Step 2.2, the static shape stream includes: a CNN-based static encoder. Teacher model based on frozen DINOv2 network and respectively , The processing yields the k-th standard static shape feature under the m-th exposure window. and standard event characteristics Used to construct the alignment loss under the m-th exposure window ; Step 3: Construct a fusion module to process the standard dynamic motion feature matrix { and standard static shape feature matrix { Processing is performed to obtain embedded features. ; Step 4: Construct the recognition module, including Each feedforward network, and respectively for The process is performed to obtain R predicted identities corresponding to the standard event stream. ,in, It is the r-th predicted identity; Step 5: Perform the following steps according to the procedure described in Step 2.1. and The process is performed to obtain the dynamic motion features of the k-th positive sample under the m-th exposure window. Dynamic motion features of the k-th negative sample under the m-th exposure window ; Following the process described in step 2.2, respectively... and The processing yields the static shape features of the positive sample under the m-th exposure window. and the static shape features of the negative sample under the m-th exposure window ; Following the process described in step 3, the dynamic motion feature matrix of the positive samples { and positive sample static shape feature matrix { Processing is performed to obtain positive sample features. Meanwhile, for the negative sample dynamic motion feature matrix { and negative sample static shape feature matrix { Processing is performed to obtain negative sample features. ; Step 6: Construct the total loss function The parameters are used to train the gait recognition network, thereby obtaining the optimal gait recognition model for gait recognition of the input event stream.

[0010] The event-based gait recognition method described in this invention is characterized in that step 2.1 is performed as follows: Step 2.1.1, the lightweight pulse gating network is Each spiking neural network, and respectively for Processing is performed, and the corresponding results are obtained. A standard adaptive mixing coefficient { ,…, ,…, },in, Indicates the first A standard adaptive mixing coefficient; Step 2.1.2 Several parallel pulse experts respectively Processing is performed, and the corresponding results are obtained. A standard expert characteristic { };in, Indicates the first A standard expert characteristic; Step 2.1.3: Use equation (1) to obtain the k-th standard dynamic motion feature under the m-th exposure window. : (1).

[0011] Furthermore, step 2.2 is performed as follows: Step 2.2.1, the teacher model right The process is performed to obtain the standard teacher features under the m-th exposure window. ; Step 2.2.2, the static encoder right The process is performed to obtain the standard static shape features under the m-th exposure window. Then The input is processed in an aligned convolutional layer to obtain the standard event features under the m-th exposure window. ; Step 2.2.3: Construct the alignment loss for the m-th exposure window using equation (2). : (2).

[0012] Furthermore, step 6 is performed as follows: Step 6.1: Construct the cross-entropy loss using equation (3). : (3) In equation (3), Let r be the probability of the predicted identity, and we have: = (4) Step 6.2: Construct the triplet loss using equation (5) : (5) In equation (5), Represents Euclidean distance. The preset boundary threshold, express function; Step 6.3: Construct the total alignment loss using equation (6). : (6) In equation (6), These are the weighting coefficients for the alignment loss; Step 6.4: Construct the total loss function using equation (7). : (7).

[0013] The present invention provides an electronic device, including a memory and a processor, characterized in that the memory is used to store a program supporting the processor in performing the method described therein, and the processor is configured to execute the program stored in the memory.

[0014] The present invention discloses a computer-readable storage medium storing a computer program, characterized in that the computer program is executed by a processor to perform the steps of the method described thereon.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention proposes an event-based dual-stream gait recognition framework (EventGait), which creatively employs a strategy of separating long and short time slices. By dynamically processing short-term slices to preserve high-frequency motion details and statically processing long-term slices to obtain a stable spatial structure, this effectively solves the temporal resolution loss and motion blur problems caused by over-aggregation in existing methods. Experimental results show that this invention outperforms existing RGB and event-based methods on both synthetic datasets (SUSTech1K-E) and real-world scene datasets, especially demonstrating significant advantages in low-light (nighttime) scenes.

[0016] 2. This invention designs a Hybrid Pulse Expert Module (MoSE), which greatly enhances the model's adaptability to complex environments. Addressing the problem that a single spiking neuron cannot simultaneously process event streams of different frequencies, this invention introduces parallel pulse experts with different membrane time constants and uses a gating network to adaptively calculate the mixing coefficients. This design enables the network to dynamically adjust its perception mode according to the density and frequency of input events (corresponding to different light intensities and movement speeds), thereby extracting robust gait dynamic features even in changing environments.

[0017] 3. This invention proposes a cross-modal structure alignment strategy (CroSA), which effectively solves the shape modeling challenge caused by the sparsity of event data space. By introducing a pre-trained large visual model as a teacher network, synchronous RGB frames (only during the training phase) are used to guide the learning of the event-based static encoder. This knowledge distillation mechanism forces the event encoder to "fill in" dense, human-structure-compliant shape features from the sparse event stream, significantly improving the discriminative power of the features, and maintaining high performance without requiring RGB data during the inference phase.

[0018] 4. This invention constructs and open-sources a large-scale event-based gait recognition dataset, filling a data gap in this field. By establishing a high-fidelity synthesis pipeline from RGB video to event stream, the generated SUSTech1K-E and CCGR-Mini-E datasets cover various lighting, viewpoint, and occlusion conditions, which not only enhances the generalization training ability of the model but also provides a standardized benchmark for evaluating event-based gait recognition algorithms. Attached Figure Description

[0019] Figure 1 A flowchart of the inventive method; Figure 2 This is a diagram of the dual-stream network framework of the present invention, wherein Event Streams are event streams, Short-term Slices are short-term event slices, Long-term Slices are long-term event slices, Dynamic Motion Stream is dynamic motion stream, and Static Shape Stream is static shape stream; Figure 3 This is a detailed diagram of the dynamic motion flow in this invention, where Mixture of Spiking Experts refers to the mixed pulse experts; Figure 4 This is a detailed diagram of the static shape flow in this invention, where Static Encoder is a static encoder, Grayscale conversion is grayscale processing, and Alignment function is an alignment function. Detailed Implementation

[0020] In this embodiment, an event-based gait recognition method is proposed to overcome the shortcomings of existing event-based gait recognition methods, such as loss of dynamic information, sparse spatial features, and poor adaptability to changing environments. It models fine motion and dense shape separately using a two-stream architecture and utilizes a hybrid expert mechanism and cross-modal alignment technology to significantly improve gait recognition accuracy in complex lighting scenarios. Specifically, see [link to documentation]. Figure 1 The method is performed according to the following steps: Step 1: Use the event camera to collect standard event streams and positive sample event streams for the same real identity, as well as negative sample event streams for different real identities. In this embodiment, the SUSTech1K / SUSTech1K-E dataset, CCGR-Mini / CCGR-Mini-E dataset, and CASIA-B* / EV-CASIA-B dataset are used to train and evaluate the model, respectively. CCGR-Mini-E is synthesized from the CCGR-Mini dataset using our synthesis pipeline, SUSTech1K-E is synthesized from the SUSTech1K dataset using our synthesis pipeline, and EV-CASIA-B is a playback-based event dataset derived from the CASIA-B dataset.

[0021] The standard event stream, positive sample event stream, and negative sample event stream are respectively divided into... One exposure window; The standard events, positive sample events, and negative sample events under the m-th exposure window are respectively divided into... The intervals are divided into several segments, and the standard events, positive sample events, and negative sample events in each interval are aggregated into short-term standard event slices, short-term positive sample event slices, and short-term negative sample event slices, respectively.

[0022] Let the short-term standard event slice corresponding to the k-th interval under the m-th exposure window be denoted as . Let the slice of short-term positive sample events corresponding to the k-th interval under the m-th exposure window be denoted as Let the slice of short-term negative sample events corresponding to the k-th interval under the m-th exposure window be denoted as .

[0023] Aggregate the standard events under the m-th exposure window into the m-th long-term standard event slice. The positive sample events under the m-th exposure window are aggregated into the m-th long-term positive sample event slice. The negative sample events under the m-th exposure window are aggregated into the m-th long-term negative sample event slice. .

[0024] Obtain the RGB image captured by a regular camera corresponding to the m-th exposure window of the standard event stream and perform grayscale processing to obtain the intensity image under the m-th exposure window. .

[0025] One-hot encoding of the real identity corresponding to the standard event stream =[ ,..., ,… ], For the r-th real identity code, This indicates the total number of identities.

[0026] Step 2: Construct an EventGait dual-stream network, including a dynamic motion flow and a static shape flow. In this embodiment, the dual-stream framework is as follows: Figure 2 As shown; Step 2.1: The dynamic motion flow consists of a lightweight pulse-gated network and A team of parallel pulse experts was formed, and [the team] conducted [the following]... The process is performed to obtain the k-th standard dynamic motion feature under the m-th exposure window. In this embodiment, as Figure 3 As shown.

[0027] Step 2.1.1: The lightweight pulse gating network is... Each spiking neural network, and respectively for Processing is performed, and the corresponding results are obtained. A standard adaptive mixing coefficient { ,…, ,…,},in, Indicates the first A standard adaptive mixing coefficient; Step 2.1.2 Several parallel pulse experts respectively Processing is performed, and the corresponding results are obtained. A standard expert characteristic { };in, Indicates the first A standard expert characteristic.

[0028] Step 2.1.3: Use equation (1) to obtain the k-th standard dynamic motion feature under the m-th exposure window. : (1) Step 2.2, in this embodiment, as follows: Figure 4 As shown, the static shape stream includes: a CNN-based static encoder. Teacher model based on frozen DINOv2 network and respectively , The processing yields the k-th standard static shape feature under the m-th exposure window. and standard event characteristics Used to construct the alignment loss under the m-th exposure window .

[0029] Step 2.2.1, Teacher Model right The process is performed to obtain the standard teacher features under the m-th exposure window. ; Step 2.2.2, Static Encoder right The process is performed to obtain the standard static shape features under the m-th exposure window. Then The input is processed in an aligned convolutional layer to obtain the standard event features under the m-th exposure window. .

[0030] Step 2.2.3: Construct the alignment loss for the m-th exposure window using equation (2). : (2) Step 3: Construct a fusion module to process the standard dynamic motion feature matrix { and standard static shape feature matrix { Processing is performed to obtain embedded features. .

[0031] Step 4: Construct the recognition module, including Each feedforward network, and respectively for The process yields R predicted identities corresponding to the standard event stream. ,in, It is the r-th predicted identity; Step 5: Follow the process in Step 2.1 to perform the following steps respectively. and The process is performed to obtain the dynamic motion features of the k-th positive sample under the m-th exposure window. Dynamic motion features of the k-th negative sample under the m-th exposure window .

[0032] Follow the process in step 2.2 to respectively and The processing yields the static shape features of the positive sample under the m-th exposure window. and the static shape features of the negative sample under the m-th exposure window ; Following step 3, the dynamic motion feature matrix of the positive samples { and positive sample static shape feature matrix { Processing is performed to obtain positive sample features. Meanwhile, for the negative sample dynamic motion feature matrix { and negative sample static shape feature matrix { Processing is performed to obtain negative sample features. .

[0033] Step 6: Construct the total loss function The parameters are used to train the gait recognition network to obtain the optimal gait recognition model, which is used to perform gait recognition on the input event stream. In this embodiment, a gait recognition model with weight decay and an initial learning rate of is used. The SGD optimizer updates the parameters of the gait recognition network, with an initial learning rate. Set the value to 0.1, and the weight decay to 0.0005; Step 6.1: Construct the cross-entropy loss using equation (3). : (3) In equation (3), Let r be the probability of the predicted identity, and we have: = (4) Step 6.2: Construct the triplet loss using equation (5) : (5) In equation (5), Represents Euclidean distance. The preset boundary threshold, express function.

[0034] Step 6.3: Construct the total alignment loss using equation (6). : (6) In equation (6), These are the weighting coefficients for the alignment loss.

[0035] Step 6.4: Construct the total loss function using equation (7). : (7).

[0036] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described method, and the processor is configured to execute the program stored in the memory.

[0037] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

[0038] Example To verify the effectiveness of the method of the present invention, the SUSTech1K / SUSTech1K-E dataset, CCGR-Mini / CCGR-Mini-E dataset, and CASIA-B* / EV-CASIA-B dataset were selected for training and testing in this embodiment.

[0039] The method was trained on SUSTech1K / SUSTech1K-E, CCGR-Mini / CCGR-Mini-E and CASIA-B* / EV-CASIA-B respectively, and evaluated on the test sets of each dataset respectively.

[0040] In this invention, the first-place hit rate (Rank-1), mean accuracy (mAP), and mean negative penalty (mINP) under normal (NM), changing clothes (CL), night (NT) conditions, and overall conditions are used as evaluation indicators.

[0041] In this embodiment, six methods with different input modalities were selected for comparison with the method of the present invention on the SUSTech1K dataset. The selected methods are LidarGait++, HMRNet, GaitBase, DeepGaitV2, GaitLLM-10, and EVGait. Six methods with different input modalities were selected for comparison with the method of the present invention on the CCGR-Mini and CASIA-B* datasets. The selected methods are GaitSet, GaitPart, GaitBase, DeepGaitV2, GPGait, and GaitGraph2. Two methods with different input modalities were selected for comparison with the method of the present invention under different lighting conditions on the SUSTech1K-E dataset. The selected methods are GaitSet and GaitBase.

[0042] The GaitBase method shown in Tables 1, 2, and 3 refers to a high-performance, general-purpose baseline gait recognition method based on RGB silhouettes.

[0043] The DeepGaitV2 shown in Tables 1 and 2 refers to an improved deep learning gait recognition method.

[0044] The GaitSet shown in Tables 2 and 3 refers to a deep learning gait recognition method based on RGB silhouette sets.

[0045] In Table 1, LidarGait++ refers to a 3D gait recognition method based on lidar point clouds, HMRNet refers to a free-environment gait recognition method based on a single lidar, GaitLLM-10 refers to a gait recognition method that combines the sequence modeling capabilities of a large language model (LLM), and EVGait refers to an early event-stream-based gait recognition method.

[0046] Table 2 shows GaitPart, a gait recognition method based on local features; GPGait, a gait recognition method based on human skeleton pose estimation; and GaitGraph2, a skeleton gait recognition method based on graph neural networks (GCN).

[0047] The experimental results are shown in Tables 1, 2, and 3. Table 1. Comparison of recognition accuracy of the method of this invention with existing mainstream methods on the SUSTech1K dataset. Table 2. Comparison of recognition accuracy of the method of the present invention with existing mainstream methods on the CCGR-Mini and CASIA-B* datasets. Table 3. Comparison of recognition accuracy of the method of this invention with existing mainstream methods on the SUSTech1K-E dataset under different lighting conditions. Experimental results show that on various datasets such as SUSTech1K-E, CCGR-Mini-E, and the real-world DVS128-Gait dataset, the proposed method outperforms existing methods based on RGB cameras, LiDAR, and other event streams, thus proving the feasibility of the proposed method. Experiments demonstrate that the proposed dual-stream framework effectively combines the high-frequency dynamic characteristics and static structural information of the event stream. Through the Hybrid Pulse Expert Module (MoSE), it achieves adaptive perception of multi-scale dynamic changes under different lighting and motion speeds. Furthermore, through Cross-Modal Structure Alignment (CroSA), it effectively solves the challenge of extracting dense human shape features from sparse event streams, thereby achieving excellent gait recognition performance in complex, uncontrolled scenes (especially in low-light environments).

Claims

1. An event-based gait recognition method, characterized in that, The procedure is as follows: Step 1: Use the event camera to collect standard event streams and positive sample event streams for the same real identity, as well as negative sample event streams for different real identities; The standard event stream, positive sample event stream, and negative sample event stream are respectively divided into... One exposure window; The standard events, positive sample events, and negative sample events under the m-th exposure window are respectively divided into... The standard events, positive sample events, and negative sample events in each interval are aggregated into short-term standard event slices, short-term positive sample event slices, and short-term negative sample event slices, respectively. Let the short-term standard event slice corresponding to the k-th interval under the m-th exposure window be denoted as . Let the slice of short-term positive sample events corresponding to the k-th interval under the m-th exposure window be denoted as Let the slice of short-term negative sample events corresponding to the k-th interval under the m-th exposure window be denoted as ; Aggregate the standard events under the m-th exposure window into the m-th long-term standard event slice. The positive sample events under the m-th exposure window are aggregated into the m-th long-term positive sample event slice. The negative sample events under the m-th exposure window are aggregated into the m-th long-term negative sample event slice. ; Obtain the RGB image captured by a regular camera corresponding to the m-th exposure window of the standard event stream and perform grayscale processing to obtain the intensity image under the m-th exposure window. ; One-hot encoding of the real identity corresponding to the standard event stream =[ ,..., ,… ], For the r-th real identity code, Indicates the total number of identities; Step 2: Construct an EventGait dual-stream network, including: dynamic motion flow and static shape flow; Step 2.1, the dynamic motion flow consists of a lightweight pulse-gated network and A team of parallel pulse experts was formed, and [the team] conducted [the following]... The process is performed to obtain the k-th standard dynamic motion feature under the m-th exposure window. ; Step 2.2, the static shape stream includes: a CNN-based static encoder. Teacher model based on frozen DINOv2 network and respectively , The processing yields the k-th standard static shape feature under the m-th exposure window. and standard event characteristics Used to construct the alignment loss under the m-th exposure window ; Step 3: Construct a fusion module to process the standard dynamic motion feature matrix { and standard static shape feature matrix { Processing is performed to obtain embedded features. ; Step 4: Construct the recognition module, including Each feedforward network, and respectively for The process is performed to obtain R predicted identities corresponding to the standard event stream. ,in, It is the r-th predicted identity; Step 5: Perform the following steps according to the procedure described in Step 2.

1. and The process is performed to obtain the dynamic motion features of the k-th positive sample under the m-th exposure window. Dynamic motion features of the k-th negative sample under the m-th exposure window ; Following the process described in step 2.2, respectively... and The processing yields the static shape features of the positive sample under the m-th exposure window. and the static shape features of the negative sample under the m-th exposure window ; Following the process described in step 3, the dynamic motion feature matrix of the positive samples { and positive sample static shape feature matrix { Processing is performed to obtain positive sample features. Meanwhile, for the negative sample dynamic motion feature matrix { and negative sample static shape feature matrix { Processing is performed to obtain negative sample features. ; Step 6: Construct the total loss function The parameters are used to train the gait recognition network, thereby obtaining the optimal gait recognition model for gait recognition of the input event stream.

2. The event-based gait recognition method according to claim 1, characterized in that, Step 2.1 is performed as follows: Step 2.1.1, the lightweight pulse gating network is Each spiking neural network, and respectively for Processing is performed, and the corresponding results are obtained. A standard adaptive mixing coefficient { ,…, ,…, },in, Indicates the first A standard adaptive mixing coefficient; Step 2.1.2 Several parallel pulse experts respectively Processing is performed, and the corresponding results are obtained. A standard expert characteristic { };in, Indicates the first A standard expert characteristic; Step 2.1.3: Use equation (1) to obtain the k-th standard dynamic motion feature under the m-th exposure window. : (1)。 3. The event-based gait recognition method according to claim 2, characterized in that, Step 2.2 is performed as follows: Step 2.2.1, the teacher model right The process is performed to obtain the standard teacher features under the m-th exposure window. ； Step 2.2.2, the static encoder right The process is performed to obtain the standard static shape features under the m-th exposure window. Then The input is processed in an aligned convolutional layer to obtain the standard event features under the m-th exposure window. ; Step 2.2.3: Construct the alignment loss for the m-th exposure window using equation (2). : (2)。 4. The event-based gait recognition method according to claim 3, characterized in that, Step 6 is performed as follows: Step 6.1: Construct the cross-entropy loss using equation (3). : (3) In equation (3), Let r be the probability of the predicted identity, and we have: = (4) Step 6.2: Construct the triplet loss using equation (5) : (5) In equation (5), Represents Euclidean distance. The preset boundary threshold, express function; Step 6.3: Construct the total alignment loss using equation (6). : (6) In equation (6), These are the weighting coefficients for the alignment loss; Step 6.4: Construct the total loss function using equation (7). : (7)。 5. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports a processor in executing the method of any one of claims 1-4, the processor being configured to execute the program stored in the memory.

6. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program is executed by the processor to perform the steps of the method according to any one of claims 1-4.

Images