Multi-person gait recognition method based on time sequence silhouette completion, medium and equipment

Abstract

The application provides a multi-person gait recognition method based on time sequence silhouette completion, a medium and equipment; wherein the method is: acquiring a plurality of walking video sequences of pedestrians; pre-processing the plurality of video sequences of pedestrians to obtain a sequence of silhouette graphs; inputting the sequence of silhouette graphs into a lightweight time sequence silhouette completion network based on double discriminator supervision to obtain a completed sequence of silhouette graphs; inputting the completed sequence of silhouette graphs into a gait feature extraction network to obtain multi-person gait feature encoding; performing feature matching on the multi-person gait feature encoding and registered library personnel gait features to obtain a final gait recognition result. The method adds a step of completing the sequence of silhouette graphs in the gait recognition process, and the occluded silhouette graph sequence is completed in a completion mode before the sequence of silhouette graphs is recognized, so that the occluded body parts formed in the multi-person scene can be automatically completed according to the information of the front and rear frames of the sequence of silhouette graphs, and the quality of the sequence of silhouette graphs is improved.

Description

Technical Field

[0001] This invention relates to the field of biometric recognition technology, and more specifically, to a method, medium, and device for multi-person gait recognition based on temporal silhouette completion. Background Technology

[0002] Gait recognition is an emerging biometric identification technology. The reason gait is considered a biometric identification technology stems from the uniqueness of each individual's gait. From an anatomical perspective, the physical basis for gait uniqueness lies in the differences in each person's physiological structure: different leg bone lengths, different muscle strengths, different center of gravity heights, and different motor nerve sensitivities all contribute to the uniqueness of gait. Gait recognition aims to identify individuals through their walking posture.

[0003] Gait recognition, as a mid-to-long-range, non-intrusive recognition method, has broad application prospects and practical value in fields such as security monitoring, privacy protection, and access control in key locations. However, gait recognition is currently affected by many factors, such as dynamic and static occlusion, clothing occlusion, and occlusion by objects carried in real-world multi-person gait recognition scenarios. These factors lead to lower gait recognition accuracy in occluded situations in real-world scenarios. Therefore, gait recognition in real-world multi-person scenarios is a problem that urgently needs to be solved.

[0004] Currently, the mainstream methods for multi-person gait recognition include: 1. Sensor-based methods: Unlike traditional gait recognition which uses RGB cameras to capture images of multiple pedestrians, sensor-based multi-person gait recognition algorithms use millimeter-wave, event cameras, depth cameras, and other types of sensors to detect multiple target pedestrians and acquire multi-target pedestrian walking data unaffected by occlusion. For example, the Chinese invention patent "A Millimeter-Wave Based Multi-Person Gait Recognition Method and Device" (Publication No.: CN111368635A) adopts this technical approach. Sensor-based methods can utilize the advantages of different sensors to avoid occlusion problems between multiple pedestrians at the data level, acquiring relatively complete pedestrian gait data. However, this type of method relies on expensive sensor equipment, and in normal, unobstructed scenarios, the feature information contained in the gait data acquired by millimeter-wave, event camera, and depth camera sensors is less than that in image frames acquired using RGB cameras. 2. Recognition methods based on human 3D information: Multiple cameras are used to model the human bodies of multiple target pedestrians, acquiring 3D gait information of different pedestrians. Then, graph convolution and other structures are used to extract gait features, achieving recognition even in occluded scenarios. For example, the Chinese invention patent application "A Method for Gait Recognition of Multiple People Based on Human Skeleton Heatmap" (Publication No.: CN114187665A) utilizes 3D human skeleton heatmap data for gait recognition. The same application uses OpenPose3D to extract 3D pose data of pedestrians from videos and employs the 3D gait recognition network model TSLN-PoseGait for gait feature extraction. This type of method can utilize cameras from multiple perspectives to construct 3D human gait information. By complementing information between different views, the influence of mutual occlusion among multiple people is reduced, resulting in relatively complete gait information. III. Methods Based on Robust Feature Extraction Networks: Optimizing the structure of the gait feature extraction network improves the feature robustness of the gait recognition model in occluded scenarios. For example, the Chinese invention patent application "A Multi-Person Gait Recognition Method Based on Spatiotemporal Hybrid Features" (Publication No.: CN112861605A) proposes a multi-person gait recognition method based on spatiotemporal hybrid features. This method utilizes a pseudo-3D residual network to extract features from the upper and lower body contours of pedestrians, and then stitches these features together during horizontal pyramid pooling to obtain occlusion-robust gait features. This type of method avoids the requirement for multiple types and multiple target sensor devices, improving upon existing gait recognition networks. However, the upper limit of this method depends on the quality of the training data, and the improvement effect is quite limited. Summary of the Invention

[0005] To overcome the shortcomings and deficiencies in the existing technology, the purpose of this invention is to provide a method, medium and device for multi-person gait recognition based on time-series silhouette completion. This method adds a silhouette image sequence completion step to the gait recognition process. Before the silhouette image sequence is recognized, the occluded silhouette image sequence is completed using a completion method. It can automatically complete the occluded human body parts caused by multi-person scenes based on the information of the previous and next frames of the silhouette image sequence, thereby improving the quality of the silhouette image sequence.

[0006] To achieve the above objectives, the present invention provides a multi-person gait recognition method based on temporal silhouette completion, comprising the following steps:

[0007] S1. Obtain video sequences of multiple pedestrians walking, and use pedestrian detection algorithms and multi-target tracking algorithms to locate pedestrians;

[0008] S2. The multi-pedestrian video sequence obtained in step S1 is segmented to obtain a black and white outline sequence for each pedestrian, and gait alignment is performed to obtain a silhouette sequence of uniform size.

[0009] S3. Input the silhouette image sequence into a lightweight temporal silhouette completion network supervised by dual discriminators to obtain the completed silhouette image sequence.

[0010] S4. Input the completed silhouette image sequence into the gait feature extraction network to obtain the gait feature encoding of multiple pedestrians;

[0011] S5. Perform feature matching between the gait feature encoding of multiple pedestrians and the gait features of people in the registered database to obtain the final gait recognition result.

[0012] Preferably, in S3, the lightweight temporal silhouette completion network based on dual discriminator supervision includes a completion network ontology, a authenticity discriminator, and an identity discriminator;

[0013] The completion network ontology is an Encoder-Decoder structure, using BN regularization and LeakyReLu as the activation function. The completion network ontology includes an encoder and a decoder. The encoder contains N convolutional layers for downsampling the silhouette image sequence to obtain compressed gait sequence features. The decoder contains N+1 deconvolutional layers to decode the compressed gait sequence features to obtain the completed silhouette image sequence.

[0014] Preferably, the lightweight temporal silhouette completion network based on dual discriminator supervision refers to a trained lightweight temporal silhouette completion network based on dual discriminator supervision. During the training process of the lightweight temporal silhouette completion network based on dual discriminator supervision, a real / fake discriminator and an identity discriminator are added to the network to optimize the ontology parameters W of the completion network. G ;

[0015] The authenticity discriminator consists of five 3D convolutional layers, using LayerNorm regularization and LeakyReLu as the activation function. The input to the authenticity discriminator is the sequence of completed silhouette images. The authenticity discriminator is responsible for judging the completion effect of the completion network body to obtain a high-quality sequence of completed silhouette images.

[0016] The identity discriminator consists of five 2D convolutional layers, using BN regularization and LeakyReLu as the activation function. The input to the identity discriminator is the gait energy map after compression and stitching of the completed silhouette image sequence and the single-person walking video sequence in the time dimension. The identity discriminator is responsible for judging the consistency of gait identity features of the completed network ontology before and after the completed silhouette image sequence, so as to prevent the loss of pedestrian gait information in the completed silhouette image sequence.

[0017] Preferably, the lightweight temporal silhouette completion network training method based on dual discriminator supervision is as follows:

[0018] Y1. A dataset of silhouette sequences with occlusion is generated by synthesizing two or more single-person walking video sequences. Specifically, two or more single-person walking video sequences are collected; gait silhouette sequences corresponding to each single-person walking video sequence are obtained based on a static background subtraction method, and gait alignment is performed; the sequences are aggregated in the image space, treating each gait silhouette sequence as a main pedestrian, and mutually occluding each other at the pixel level to form a dataset of silhouette sequences with occlusion.

[0019] Y2, Initialize and complete the network ontology G, and the authenticity detector D. R Identity discrimination network D I The corresponding parameters are W G ,

[0020] Sampling of occluded silhouette sequence x and undisturbed ground truth silhouette sequence x from a dataset of occluded silhouette sequence datasets. The silhouette sequence has the shape (BS, L, W, H); where BS represents the batch size, L represents the sequence length, W represents the width of the silhouettes in the sequence, and H represents the height. A random sample of the positive example silhouette sequence x... pos And negative example silhouette sequence x neg ;

[0021] Set training parameters: Number of training iterations P n Number of training iterations I n LR (Learning Rate) for Complete Network Ontology G True / False Discriminant Learning Rate (LR) r Identity discriminator learning rate LR I The frequency of alternating iterations between the network ontology and the authenticity and identity discriminators is K, the loss weight parameters α and β, the gradient penalty term coefficient λ, the Adam optimizer parameters, and the random sampled values ​​θ∈[0,1].

[0022] Y3, Loop P n Each traversal contains I. n Training iterations;

[0023] In each training iteration, the loss value of the true / false discriminator is calculated.

[0024] L WGAN =D R (G(x))-D R (x)

[0025]

[0026]

[0027] Among them, L WGAN For Wasserstein distance loss; L GP The loss is a gradient penalty. Indicates the network backpropagation gradient;

[0028] Update the parameters of the true / false discriminant using a gradient optimization algorithm.

[0029]

[0030] In each training iteration, the loss value of the identity discriminator is calculated.

[0031]

[0032]

[0033]

[0034] in, For the positive example silhouette image sequence x pos Image compressed in the time dimension; For negative example silhouette image sequence x negImage compressed in the time dimension; This is a time-compressed image of a sequence of silhouette images sampled in a training set; torch.mean() is the mean operation; dim is the dimension corresponding to the operation; bceloss() is the cross-entropy loss function; torch.oneslike() is a tensor with a value of 1;

[0035] torch.zeroslike() is a tensor with a value of 0; logit s1 D I exist Output of a stitched image input; logit s2 D I exist Output of a stitched image input; logit d D I exist Output from a stitched image input;

[0036] Update the identity discriminator parameters using a gradient optimization algorithm.

[0037]

[0038] After every K training iterations, calculate the loss value L for completing the network ontology. G :

[0039]

[0040]

[0041]

[0042] L ID =bceloss(logit d torch.oneslike(logit d ))

[0043] L G =αL IMG +L WGAN +βL ID

[0044] Among them, L IMG For image reconstruction loss; mseloss() is the mean loss; L WGAN Wasserstein distance loss;

[0045] The gradient optimization algorithm is used to update and complete the network ontology parameters W. G :

[0046]

[0047] Y4, in P n After a series of traversals, the training of the network ontology completion, authenticity discriminator, and identity discriminator is completed.

[0048] Preferably, Y1 refers to acquiring two or more single-person walking video sequences; obtaining gait silhouette image sequences corresponding to each single-person walking video sequence based on a static background subtraction method, and performing gait alignment operations; aggregating in the image space, treating each gait silhouette image sequence as a main pedestrian, and mutually occluding each other at the pixel level to form a dataset of silhouette image sequences with occlusion.

[0049] Preferably, in step S1, the pedestrian detection algorithm uses the single-stage target detection YOLOv51 model or the Faster R-CNN model; the multi-target pedestrian tracking algorithm uses the ByteTrack algorithm or the TrackFormer algorithm.

[0050] Preferably, in S4, the gait feature extraction network refers to the GaitSet gait feature extraction network based on an unordered set.

[0051] Preferably, S5 refers to: performing distance matching between the gait feature encoding of multiple pedestrians and the gait features of people in the registered database, using Euclidean distance as the distance metric in high-dimensional space, to obtain the final gait recognition result.

[0052] A readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform the aforementioned multi-person gait recognition method based on temporal silhouette completion.

[0053] A computer device includes a processor and a memory for storing a processor-executable program, wherein when the processor executes the program stored in the memory, it implements the above-described method for multi-person gait recognition based on temporal silhouette completion.

[0054] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0055] Compared to existing sensor-based multi-person gait recognition methods, this invention can use relatively inexpensive surveillance cameras to capture images of multiple pedestrians without relying on expensive sensor equipment. Furthermore, in normal multi-person unobstructed scenarios, this invention uses an RGB camera to acquire pedestrian gait frames, which contain more information than gait data acquired by sensors based on millimeter-wave, event cameras, depth cameras, and other types of sensors.

[0056] Compared with existing multi-person gait recognition methods based on human 3D information, this invention only requires pedestrian information obtained from a single viewpoint to perform occlusion completion and quality improvement based on the relationship between gait frames before and after the gait. It does not require cameras from multiple viewpoints to construct 3D human gait information, thus reducing scene requirements and equipment costs.

[0057] Compared to multi-person gait methods based on robust feature extraction networks, this invention performs occlusion completion and quality improvement on silhouette images during the preprocessing stage, eliminating the need to retrain the gait feature extraction network, thus reducing the quality requirements of network training data and the computational power required to train large-scale networks. Attached Figure Description

[0058] Figure 1 This is a flowchart illustrating the multi-person gait recognition method based on temporal silhouette completion of the present invention.

[0059] Figure 2 This is a schematic diagram of the structure of the network ontology completed in this invention;

[0060] Figure 3 This is a schematic diagram of the authenticity detector of the present invention;

[0061] Figure 4 This is a schematic diagram of the identity discriminator of the present invention. Detailed Implementation

[0062] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0063] Example 1

[0064] This embodiment presents a multi-person gait recognition method based on temporal silhouette completion, such as... Figure 1 As shown, it includes the following steps:

[0065] S1. Acquire video sequences of multiple pedestrians and locate pedestrians using pedestrian detection and multi-object tracking algorithms. The pedestrian detection algorithm uses the single-stage object detection model YOLOv5l or Faster R-CNN; the multi-object pedestrian tracking algorithm uses the ByteTrack algorithm or TrackFormer algorithm.

[0066] S2. The multi-pedestrian video sequence obtained in step S1 is segmented to obtain a black and white outline sequence for each pedestrian, and gait alignment is performed to obtain a silhouette sequence of uniform size.

[0067] S3. Input the silhouette image sequence into a lightweight temporal silhouette completion network based on dual discriminator supervision to obtain the completed silhouette image sequence.

[0068] A lightweight temporal silhouette completion network based on dual discriminator supervision includes a completion network ontology.

[0069] like Figure 2 As shown, the completion network ontology is an Encoder-Decoder structure. The completion network ontology uses BN regularization and LeakyReLu as the activation function. The completion network ontology includes an encoder and a decoder. The encoder contains N convolutional layers for downsampling the silhouette image sequence to obtain compressed gait sequence features. The decoder contains N+1 deconvolutional layers for decoding the compressed gait sequence features to obtain the completed silhouette image sequence.

[0070] S4. Input the completed silhouette sequence into the gait feature extraction network to obtain the gait feature encoding of multiple pedestrians. The gait feature extraction network can be implemented using the GaitSet gait feature extraction network based on unordered sets or other gait feature extraction networks.

[0071] S5. The gait feature encoding of multiple pedestrians is matched with the gait features of people in the registration database by distance. Euclidean distance is used as the distance metric in high-dimensional space to obtain the final gait recognition result.

[0072] The lightweight temporal silhouette completion network under dual discriminator supervision refers to a trained lightweight temporal silhouette completion network under dual discriminator supervision. The training method for the lightweight temporal silhouette completion network under dual discriminator supervision is as follows:

[0073] Y1. Synthesize two or more single-person walking video sequences to form a dataset of silhouette images with occlusion.

[0074] Y2. Add a real / fake discriminator and an identity discriminator to the lightweight temporal silhouette completion network based on dual discriminator supervision;

[0075] like Figure 3 As shown, the authenticity discriminator consists of five 3D convolutional layers, using LayerNorm regularization and LeakyReLu as the activation function; the input to the authenticity discriminator is the sequence of completed silhouette images; the authenticity discriminator is responsible for judging the completion effect of the completion network body to obtain a high-quality sequence of completed silhouette images.

[0076] like Figure 4 As shown, the identity discriminator consists of five 2D convolutional layers, using BN regularization and LeakyReLu as the activation function. The input to the identity discriminator is the gait energy map after compression and stitching of the completed silhouette image sequence and the single-person walking video sequence in the time dimension. The identity discriminator is responsible for judging the consistency of gait identity features of the completed network ontology before and after the completed silhouette image sequence, so as to prevent the loss of pedestrian gait information in the completed silhouette image sequence.

[0077] Initialize and complete the network ontology G and the authenticity discriminator DR Identity discrimination network D I The corresponding parameters are W G ,

[0078] Sampling of occluded silhouette sequence x and undisturbed ground truth silhouette sequence x from a dataset of occluded silhouette sequence datasets. The silhouette sequence has the shape (BS, L, W, H); where BS represents the batch size, L represents the sequence length, W represents the width of the silhouettes in the sequence, and H represents the height. A random sample of the positive example silhouette sequence x... Pos And negative example silhouette sequence x neg All of the above sequences have been cropped to 32 frames;

[0079] Set training parameters: Number of training iterations P n Number of training iterations I n LR (Learning Rate) for Complete Network Ontology G True / False Discriminant Learning Rate (LR) r Identity discriminator learning rate LR I The frequency of alternating iterations between the network ontology and the authenticity and identity discriminators is K; the loss weight parameters α and β; the gradient penalty term coefficient λ; the Adam optimizer parameters; and the random sampled values ​​θ∈[0,1].

[0080] Y3, Loop P n Each traversal contains I. n Training iterations;

[0081] In each training iteration, the loss value of the true / false discriminator is calculated.

[0082] L WGAN =D R (G(x))-D R (x)

[0083]

[0084]

[0085] Among them, L WGAN For Wasserstein distance loss; L GP The loss is a gradient penalty. Indicates the network backpropagation gradient;

[0086] Update the parameters of the true / false discriminant using a gradient optimization algorithm.

[0087]

[0088] In each training iteration, the loss value of the identity discriminator is calculated.

[0089]

[0090]

[0091]

[0092] in, For the positive example silhouette image sequence x pos Image compressed in the time dimension; For negative example silhouette image sequence x neg Image compressed in the time dimension; This is a time-compressed image of a sequence of silhouette images sampled in a training set; torch.mean() is the mean operation; dim is the dimension corresponding to the operation; bceloss() is the cross-entropy loss function; torch.oneslike() is a tensor with a value of 1;

[0093] torch.zeroslike() is a tensor with a value of 0; logit s1 D I exist Output of a stitched image input; logit s2 D I exist Output of a stitched image input; logit d D I exist Output from a stitched image input;

[0094] Update the identity discriminator parameters using a gradient optimization algorithm.

[0095]

[0096] After every K training iterations, calculate the loss value L for completing the network ontology. G :

[0097]

[0098]

[0099]

[0100] L ID =bceloss(logit d torch.oneslike(logitd )

[0101] L G =αL IMG +L WGAN +βL ID

[0102] Among them, L IMG For image reconstruction loss; mseloss() is the mean loss; L WGAN Wasserstein distance loss;

[0103] The gradient optimization algorithm is used to update and complete the network ontology parameters W. G :

[0104]

[0105] Y4, in P n After a series of traversals, the training of the network ontology completion, authenticity discriminator, and identity discriminator is completed.

[0106] The gait silhouette image sequence completion model is trained using a dual discriminator-supervised approach. It preserves the gait features of the original pedestrian sequence while restoring image occlusion, maintaining consistency in gait features before and after completion, and effectively improving gait recognition performance in multi-person scenarios.

[0107] Example 2

[0108] This embodiment provides a readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform the multi-person gait recognition method based on temporal silhouette completion as described in Embodiment 1.

[0109] Example 3

[0110] This embodiment discloses a computer device, including a processor and a memory for storing processor-executable programs. When the processor executes the program stored in the memory, it implements the multi-person gait recognition method based on temporal silhouette completion described in Embodiment 1.

[0111] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A multi-person gait recognition method based on time-series silhouette completion, characterized in that: It comprises the following steps: S1, obtaining multiple pedestrian walking video sequences, using a pedestrian detection algorithm and a multi-target tracking algorithm to locate pedestrians; S2, segmenting the multiple pedestrian video sequences obtained in step S1 to obtain a black and white contour graph sequence corresponding to each pedestrian, and performing gait alignment operation to obtain a uniform size silhouette graph sequence; S3, inputting the silhouette graph sequence into a lightweight time sequence silhouette completion network based on double discriminator supervision to obtain a completed silhouette graph sequence; S4, inputting the completed silhouette graph sequence into a gait feature extraction network to obtain multiple pedestrian gait feature encodings; S5, performing feature matching on the multiple pedestrian gait feature encodings and the registered library personnel gait features to obtain the final gait recognition result; In S3, the lightweight time sequence silhouette completion network based on double discriminator supervision comprises a completion network body; The completion network body is an Encoder-Decoder coding structure, and BN regularization means and LeakyReLu are used as the activation function in the completion network body; the completion network body comprises an encoder and a decoder; the encoder comprises N convolutional layers for downsampling operation of the silhouette graph sequence to obtain compressed gait sequence features; the decoder comprises N+1 deconvolutional layers for decoding the compressed gait sequence features to obtain the completed silhouette graph sequence; The light-weighted temporal silhouette completion network based on double discriminator supervision refers to a trained light-weighted temporal silhouette completion network based on double discriminator supervision; during training of the light-weighted temporal silhouette completion network based on double discriminator supervision, a true-false discriminator and an identity discriminator are added to the light-weighted temporal silhouette completion network based on double discriminator supervision to optimize the parameters W of the completion network body G ; The authenticity discriminator is composed of five 3D convolutional layers, uses LayerNorm regularization means, and LeakyReLu as the activation function; the input of the authenticity discriminator is the completed silhouette graph sequence; the authenticity discriminator is responsible for discriminating the completion effect of the completion network body to obtain a high image quality completed silhouette graph sequence; The identity discriminator is composed of five 2D convolutional layers, uses BN regularization means, and LeakyReLu as the activation function; the input of the identity discriminator is the completed silhouette graph sequence and the gait energy graph obtained by compressing and splicing the single person walking video sequence in the time dimension; the identity discriminator is responsible for discriminating the consistency of the gait identity features before and after the completion of the silhouette graph sequence by the completion network body to prevent the completed silhouette graph sequence from losing pedestrian gait information.

2. The multi-person gait recognition method based on temporal silhouette completion according to claim 1, characterized in that: The training method of the lightweight time sequence silhouette completion network based on double discriminator supervision is: Y1, synthesizing two or more single person walking video sequences to form a silhouette graph sequence dataset with occlusion; Y2, initialize the completion network ontology G, the authenticity discriminator D R , identity discriminator network D I , the corresponding parameters are sampling occluded silhouette sequence x and unoccluded ground truth silhouette sequence x in occluded silhouette sequence dataset with occlusion shape of silhouette sequence is (BS, L, W, H); where BS represents data batch size, L represents silhouette sequence length, W represents silhouette width in silhouette sequence, H represents silhouette height; randomly sampling positive example silhouette sequence x pos and negative example silhouette sequence x neg ; Set training parameters: training traversal times P n , training iteration times I n , learning rate of the completion network LR G , learning rate of the authenticity discriminator LR r , learning rate of the identity discriminator LR I , the alternating iteration frequency K of the completion network and the authenticity discriminator and the identity discriminator, the loss weight parameters α and β, the gradient penalty term coefficient λ, the Adam optimizer parameters, the random sampling value θ ∈ [0, 1]; Y3, cycle P n one pass, each pass containing I n training iterations; In each training iteration, the loss value of the authenticity discriminator is calculated L WGAN = D R (G(x)) - D R (x) where L WGAN is the Wasserstein distance loss; L GP is the gradient penalty loss; denotes the network backpropagated gradient; updating the authenticity discriminator parameters using a gradient optimization algorithm In each training iteration, a loss value for the identity discriminator is computed wherein, positive example silhouette image sequence x pos images compressed in time dimension; negative example silhouette image sequence x neg images compressed in time dimension; images compressed in time dimension for silhouette image sequence sampled in training set; torch.mean() is mean operation; dim is dimension corresponding to operation; bceloss() is cross-entropy loss function; torch.oneslike() is tensor with value 1; torch.zeroslike() is tensor with value 0; logit s1 D I In output under concatenated image input; logit s2 D I In output under concatenated image input; logit d D I In output under concatenated image input; updating the identity discriminator parameters using a gradient optimization algorithm After every K training iterations, the loss value L of the completion network ontology is calculated G : L ID = bce_loss(logit d , torch.ones_like(logit d )) L G = aL IMG + L WGAN + βL ID wherein L IMG is the image reconstruction loss; mseloss() is the mean loss; L WGAN is the Wasserstein distance loss; updating the parameters of the completion network W using a gradient optimization algorithm G : Y4, in P n The training of the complete network ontology, the authenticity discriminator, and the identity discriminator is completed after one traversal.

3. The multi-person gait recognition method based on temporal silhouette completion according to claim 2, characterized in that: Y1 refers to collecting two or more single person walking video sequences; obtaining the gait silhouette graph sequence corresponding to each single person walking video sequence based on the static background subtraction method and performing gait alignment operation; aggregating in the image space, regarding each gait silhouette graph sequence as a main pedestrian, and mutually occluding each other at the pixel level to form a silhouette graph sequence dataset with occlusion.

4. The multi-person gait recognition method based on temporal silhouette completion according to claim 1, characterized in that: In S1, the pedestrian detection algorithm uses a single-stage target detection YOLOv5l model or a Faster R-CNN model; the multi-target pedestrian tracking algorithm uses a ByteTrack algorithm or a TrackFormer algorithm.

5. The multi-person gait recognition method based on temporal silhouette completion according to claim 1, characterized in that: In S4, the gait feature extraction network refers to a GaitSet gait feature extraction network based on unordered sets.

6. The multi-person gait recognition method based on temporal silhouette completion according to claim 1, characterized in that: The S5 refers to: matching the multiple-row person gait feature codes with the registered library person gait features in distance, using the Euclidean distance as the distance measurement in the high-dimensional space, and obtaining the final gait recognition result.

7. A readable storage medium characterized by, The storage medium stores a computer program, and the computer program causes the processor to execute the multi-person gait recognition method based on time sequence silhouette completion in any one of claims 1-6 when executed by the processor.

8. A computer device comprising a processor and a memory for storing a processor executable program, characterized in that, The processor executes the program stored in the memory, and the multi-person gait recognition method based on time sequence silhouette completion in any one of claims 1-6 is implemented.

Images