A gait recognition method, system, electronic device and storage medium

Abstract

The application discloses a gait recognition method and system, electronic equipment and a storage medium. The method comprises the following steps: acquiring a to-be-detected gait sequence and at least two candidate gait sequences; extracting a first classification feature vector and a first angle feature vector of the to-be-detected gait sequence, and extracting a second classification feature vector and a second angle feature vector of each candidate gait sequence; calculating a first distance matrix feature corresponding to each candidate gait sequence by using the classification feature vector; calculating a second distance matrix feature corresponding to each candidate gait sequence by using the angle feature vector; fusing the distance matrix features by using a preset fusion weight to acquire a feature distance of fusion angle information corresponding to each candidate gait sequence; reordering each candidate gait sequence according to the feature distance; and identifying a target according to the reordering result and the to-be-detected gait sequence. Through the above scheme, all visual angles can be rearranged by using the acquired feature distance, and the recognition efficiency is improved.

Description

Technical Field

[0001] This application relates to the field of computer vision technology, and in particular to a gait recognition method, system, electronic device, and storage medium. Background Technology

[0002] With the increasing maturity and popularity of computer technology, gait recognition is becoming more and more important. Existing gait recognition methods typically involve three steps: pedestrian foreground segmentation, feature extraction, and gait recognition.

[0003] For the existing During the research and practice, the inventors of this application discovered that with the development of gait recognition, the accuracy requirements for perspective recognition in daily life and industry are becoming increasingly higher. In particular, for cross-perspective recognition, contour recognition is easily affected by body shape and angle. When the contour changes significantly, the cross-perspective recognition effect is poor. Existing technologies perform cross-perspective recognition by reordering one-to-one perspectives, but multiple sorting is required for multiple perspectives, resulting in low recognition efficiency. In addition, insufficient extraction of temporal information will lead to low accuracy in cross-perspective recognition. Summary of the Invention

[0004] The main technical problem addressed by this application is to provide a gait recognition method, system electronic device, and storage medium that can obtain the first classification feature vector and the first angle feature vector corresponding to the sequence to be detected through a pre-trained gait model, as well as the second classification feature vector and the second angle feature vector corresponding to several gait sequences in the gait library, thereby obtaining the corresponding first distance matrix feature and second matrix feature, and then obtaining the feature distance fused with angle information. Based on the feature distance, the viewpoint is reordered to complete the recognition of the sequence to be detected.

[0005] To address the aforementioned technical problems, this application provides a gait recognition method, comprising: acquiring a gait sequence to be detected, and at least two gait sequences from a gait database; extracting a first classification feature vector and a first angle feature vector from the gait sequence to be detected, and extracting a second classification feature vector and a second angle feature vector from each candidate gait sequence in the at least two gait sequences; calculating a first distance matrix feature corresponding to each candidate gait sequence using the first classification feature vector and the second classification feature vector of each candidate gait sequence; calculating a second distance matrix feature corresponding to each candidate gait sequence using the first angle feature vector and the second angle feature vector of each candidate gait sequence; fusing the first distance matrix feature and the second distance matrix feature using a preset fusion weight to obtain feature distances of fused angle information corresponding to each candidate gait sequence; reordering the at least two candidate gait sequences according to the feature distances between each candidate gait sequence and the gait sequence to be detected; and performing gait recognition on a target based on the reordered candidate gait sequences and the gait sequence to be detected.

[0006] In one embodiment of this application, the step of acquiring the gait sequence to be detected and at least two candidate gait sequences in the gait library includes: acquiring pedestrian gait data to be detected, and then determining the gait sequence to be detected based on the pedestrian gait data; acquiring various types of gait data in real-world scenarios, determining several candidate gait sequences based on the various types of gait data in real-world scenarios, and then forming a gait library; extracting a first classification feature vector and a first angle feature vector of the gait sequence to be detected, and extracting a second classification feature vector and a second angle feature vector of each candidate gait sequence in the at least two candidate gait sequences, including: performing feature training on the gait sequence to be detected using a pre-trained gait model, extracting the first classification feature vector and the first angle feature vector of the gait sequence to be detected, and extracting the second classification feature vector and the second angle feature vector of each candidate gait sequence in the at least two candidate gait sequences.

[0007] In one embodiment of this application, the step of extracting features from the gait sequence to be detected using a pre-trained gait model, and extracting a first classification feature vector and a first angle feature vector from the gait sequence to be detected, includes: extracting a first classification feature vector from the gait sequence to be detected using a first branch of the pre-trained gait model, and extracting a first angle feature vector from the gait sequence to be detected using a second branch of the pre-trained gait model; wherein, the first classification feature vector includes temporal features and static features, the temporal features are extracted based on the lower body trunk, and the static features are extracted based on the upper body trunk; the step of extracting a second classification feature vector and a second angle feature vector from each of the at least two candidate gait sequences includes: the second classification feature vector and the second angle feature vector are determined based on each of the at least two candidate gait sequences in the gait database.

[0008] In one embodiment of this application, the step of calculating the first distance matrix feature corresponding to each candidate gait sequence using the first classification feature vector and the second classification feature vector of each candidate gait sequence includes: obtaining the distance between the first classification feature vector and the second classification feature vector of each candidate gait sequence, and obtaining the first distance matrix feature corresponding to each candidate gait sequence based on the obtained distance; the step of calculating the second distance matrix feature corresponding to each candidate gait sequence using the first angle feature vector and the second angle feature vector of each candidate gait sequence includes: obtaining the distance between the first angle feature vector and the second angle feature vector of each candidate gait sequence, and obtaining the second distance matrix feature corresponding to each candidate gait sequence based on the obtained distance.

[0009] In one embodiment of this application, before fusing the first distance matrix feature and the second distance matrix feature using a preset fusion weight, the gait recognition method further includes: setting the preset fusion weight value range to 0-1 based on the second distance matrix feature.

[0010] In one embodiment of this application, the step of fusing the first distance matrix feature and the second distance matrix feature using the preset fusion weight to obtain the feature distance of the fusion angle information corresponding to each candidate gait sequence includes: performing a calculation operation between the first distance matrix feature and the second distance matrix feature with preset fusion weight, thereby obtaining the feature distance of the fusion angle information corresponding to each candidate gait sequence.

[0011] In one embodiment of this application, the step of reordering the at least two candidate gait sequences according to the feature distances corresponding to each candidate gait sequence and the gait sequence to be detected includes: obtaining the feature distances of the fusion angle information of each candidate gait sequence in the at least two candidate gait sequences; reordering the at least two candidate gait sequences by size according to the feature distances corresponding to each candidate gait sequence and the gait sequence to be detected, so as to determine the reordered candidate gait sequences; wherein the feature distances of the same gait angle labels are smaller than the feature distances of different gait angle labels.

[0012] In one embodiment of this application, the step of performing gait recognition on a target based on the reordered candidate gait sequence and the gait sequence to be detected includes: obtaining the reordered candidate gait sequence; comparing the feature distance between the reordered candidate gait sequence and the gait sequence to be detected; and determining the gait recognition result based on the feature distance comparison result.

[0013] In one embodiment of this application, the pre-trained gait model includes: acquiring full-view gait data of a target site, creating label data based on the full-view gait data, and determining a contour sequence based on the full-view gait data and the label data; inputting the contour sequence and extracting sequence features of the contour sequence; extracting temporal and static features of the contour sequence based on the sequence features; extracting angular features of the contour sequence based on the sequence features; training the gait sequence number recognition function of the model to be trained according to the temporal and static features; and training the gait angle recognition function of the model to be trained according to the angular features.

[0014] In one embodiment of this application, training the gait sequence recognition function of the model to be trained includes: training the gait sequence recognition function of the model to be trained using triplet loss and gait sequence classification loss; training the gait angle recognition function of the model to be trained includes: training the view angle, scene, or pitch angle of the model to be trained.

[0015] To solve the above-mentioned technical problems, another technical solution adopted in this application is: providing a gait recognition system, comprising: an acquisition module for acquiring a gait sequence to be detected and at least two candidate gait sequences in a gait database; an extraction module for extracting a first classification feature vector and a first angle feature vector of the gait sequence to be detected, and extracting a second classification feature vector and a second angle feature vector of each candidate gait sequence from the at least two candidate gait sequences; a first calculation module for calculating a first distance matrix feature corresponding to each candidate gait sequence using the first classification feature vector and the second classification feature vector of each candidate gait sequence; and a second calculation module for calculating a first distance matrix feature corresponding to each candidate gait sequence. The calculation module calculates the second distance matrix feature corresponding to each candidate gait sequence using the first angle feature vector and the second angle feature vector of each candidate gait sequence; the fusion module fuses the first distance matrix feature and the second distance matrix feature using preset fusion weights to obtain the feature distance of the fused angle information corresponding to each candidate gait sequence; the reordering module reorders the at least two candidate gait sequences according to the feature distances corresponding to each candidate gait sequence and the gait sequence to be detected; and the recognition module performs gait recognition on the target based on the reordered candidate gait sequences and the gait sequence to be detected.

[0016] In one embodiment of this application, the gait recognition system further includes: a gait model, which is pre-trained; the gait model includes at least one convolutional module, and a first branch module and a second branch module respectively connected to the at least one convolutional module; the first branch module includes a feature extraction and fusion module, a first pooling module, a first loss training module, and a second loss training module connected in sequence, and the first branch module is used to train the gait sequence to be detected for gait number recognition function; wherein, the feature extraction and fusion module is used to extract the temporal features and static features of the gait sequence to be detected; the second branch module includes a second pooling module, a third pooling module, and a third loss training module connected in sequence, and the second branch module is used to train the gait sequence to be detected for gait angle recognition function.

[0017] In one embodiment of this application, the feature extraction and fusion module includes: a first feature extraction module and a second feature extraction module; the first feature extraction module is used to extract static features, and the second feature extraction module is used to extract temporal features; wherein, the temporal features are extracted based on the lower body torso, and the static features are extracted based on the upper body torso.

[0018] To solve the above-mentioned technical problems, another technical solution adopted in this application is: to provide an electronic device, including a memory and a processor coupled to the memory, wherein the memory stores program instructions, and the processor executes the program instructions to implement the steps of the above-mentioned gait recognition method.

[0019] To solve the above-mentioned technical problems, another technical solution adopted in this application is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the gait recognition method described above.

[0020] Unlike existing technologies, the gait recognition method provided in this application includes: acquiring a sequence to be detected and at least two candidate gait sequences from a gait database; extracting a first classification feature vector and a first angle feature vector from the gait sequence to be detected, and extracting a second classification feature vector and a second angle feature vector from each of the at least two candidate gait sequences; calculating a first distance matrix feature corresponding to each candidate gait sequence using the first classification feature vector and the second classification feature vector of each candidate gait sequence; and calculating a second distance matrix feature corresponding to each candidate gait sequence using the first angle feature vector and the second angle feature vector of each candidate gait sequence; fusing the first distance matrix feature and the second distance matrix feature using preset fusion weights to obtain the feature distance of the fused angle information corresponding to each candidate gait sequence; and then... According to the feature distances between each candidate gait sequence and the gait sequence to be detected, the at least two candidate gait sequences are reordered; the target gait is identified based on the reordered candidate gait sequences and the gait sequence to be detected; that is, this application extracts the first classification feature vector and the first angle feature vector of the gait sequence to be detected, as well as the second classification feature vector and the second angle feature vector corresponding to at least two candidate gait sequences in the gait database, and then obtains the first distance matrix feature and the second distance matrix feature corresponding to each candidate gait sequence; and obtains the feature distance of the fusion angle corresponding to each candidate gait sequence based on the first distance matrix feature and the second distance matrix feature, reorders the candidate gait sequences by viewpoint according to the feature distance, and completes the identification of the gait sequence to be detected based on the reordered candidate gait sequences. All viewpoints can be reordered simultaneously, improving the identification efficiency. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating an embodiment of the gait recognition method of the present invention;

[0022] Figure 2 This is a flowchart illustrating an embodiment of step S1 of the present invention;

[0023] Figure 3 This is a flowchart illustrating an embodiment of step S2 of the present invention;

[0024] Figure 4 This is a flowchart illustrating an embodiment of step S3 of the present invention;

[0025] Figure 5 This is a flowchart illustrating another embodiment of step S3 of the present invention;

[0026] Figure 6 This is a flowchart illustrating an embodiment of step S4 of the present invention;

[0027] Figure 7 This is a flowchart illustrating an embodiment of step S5 of the present invention;

[0028] Figure 8 This is a schematic diagram of the structure of an embodiment of the gait recognition system of the present invention;

[0029] Figure 9 This is a schematic diagram of the structure of an embodiment of the electronic device of the present invention;

[0030] Figure 10 This is a schematic diagram of the structure of an embodiment of the computer-readable storage medium of the present invention. Detailed Implementation

[0031] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be particularly noted that the following embodiments are for illustrative purposes only and do not limit the scope of the invention. Similarly, the following embodiments are only some, not all, embodiments of the present invention, and all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0033] Traditional gait recognition methods only support one-to-one reordering of gait data across different perspectives. For data from multiple perspectives, multiple reordering is required, resulting in low recognition efficiency. Furthermore, the ability to extract temporal information may lead to low accuracy in cross-perspective recognition.

[0034] The applicant discovered in their research that, for data from multiple perspectives, full-view gait data can be obtained to obtain the gait sequence to be detected. By extracting the classification feature vector and angular feature vector corresponding to the gait sequence to be detected, as well as the classification feature vector and angular feature vector corresponding to the candidate gait sequences, the distance matrix features corresponding to each candidate gait sequence are obtained. Then, the feature distance of the fusion angle corresponding to each candidate gait sequence is obtained. Based on the feature distance, the perspectives are reordered to complete the recognition of the gait sequence to be detected. All perspectives can be reordered simultaneously, thereby improving the recognition efficiency.

[0035] Therefore, a gait recognition method is proposed. This method extracts the first classification feature vector and the first angle feature vector of the gait sequence to be detected, as well as the second classification feature vector and the second angle feature vector corresponding to each candidate gait sequence in at least two candidate gait sequences in the gait database. This allows the acquisition of the first distance matrix feature and the second distance matrix feature corresponding to each candidate gait sequence. Based on the first distance matrix feature and the second distance matrix feature, the method obtains the feature distance of the fusion angle corresponding to each candidate gait sequence. Then, the method reorders the viewpoints of each candidate gait sequence according to the feature distance. Finally, the method completes the recognition of the sequence to be detected based on the reordered candidate gait sequences.

[0036] Please see Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the gait recognition method of the present invention. It should be noted that if substantially the same result is obtained, the method of the present invention does not necessarily require further clarification. Figure 1 The process sequence shown is limited to the following: Figure 1 As shown, the method includes the following steps:

[0037] S1. Obtain the gait sequence to be detected, and at least two candidate gait sequences from the gait library;

[0038] Among them, the gait sequence to be detected is the contour sequence formed by pedestrian gait data that needs to be recognized across perspectives; the gait library is a database composed of several candidate gait sequences, which is used to compare and recognize the sequence to be detected. The several candidate gait sequences are obtained through pre-training.

[0039] See Figure 2 , Figure 2 This is a flowchart illustrating an embodiment of step S1, which includes:

[0040] S11. Obtain the gait data of the pedestrian to be detected, and then determine the gait sequence to be detected based on the gait data;

[0041] Among them, gait refers to the posture and behavioral characteristics of the human body when walking. The human body moves along a certain direction through a series of continuous movements of the hip, knee, ankle and toes. The gait sequence to be detected is the outline sequence of the human body when walking.

[0042] Specifically, the posture and behavioral characteristics of a human walking are acquired, and the contour sequence of a human walking is determined based on the walking posture and behavioral characteristics, which is set as the gait sequence to be detected.

[0043] S12. Obtain various types of gait data in real-world scenarios, determine several candidate gait sequences based on the various types of gait data in real-world scenarios, and then form a gait library;

[0044] The gait library consists of several candidate gait sequences determined from various gait data captured in real-world scenarios. For example, a circle of cameras is set up around the target site, and different pedestrians walk back and forth in the site to collect gait data from various perspectives. Then, several candidate gait sequences corresponding to the gait data from various perspectives are determined, thus forming the gait library. This means acquiring enough perspective data to improve the accuracy of cross-angle recognition.

[0045] S2. Using a pre-trained gait model, extract the first classification feature vector and the first angle feature vector of the gait sequence to be detected, and extract the second classification feature vector and the second angle feature vector of each candidate gait sequence from at least two candidate gait sequences.

[0046] The pre-trained gait model is trained using full-view gait data. The first classification feature vector of the gait sequence to be detected is a feature vector obtained by transforming the temporal and static features extracted from the gait sequence to be detected through the gait model. The first angle feature vector of the gait sequence to be detected is a feature vector obtained by transforming the angle features extracted from the gait sequence to be detected through the gait model. And so on, the second classification feature vector and the second angle feature vector of each candidate gait sequence in at least two candidate gait sequences are obtained.

[0047] See Figure 3 , Figure 3 This is a flowchart illustrating one embodiment of step S2, which includes:

[0048] S21. Extract the first classification feature vector of the gait sequence to be detected using the first branch of the pre-trained gait model, and extract the first angle feature vector of the gait sequence to be detected using the second branch of the pre-trained gait model; wherein, the first classification feature vector includes temporal features and static features, the temporal features are extracted based on the lower body trunk, and the static features are extracted based on the upper body trunk.

[0049] The pre-trained gait model is a neural network model. It is formed by collecting gait data from all perspectives and creating label data based on the gait data. The network model is trained based on the gait data and label data. The network model extracts the temporal dynamic features and static features of gait. At the same time, it trains an ID classifier and a gait perspective classifier. The label data includes gait angle, pitch angle, scene, and gait ID number label. The pre-trained gait model includes a convolutional module for sequence feature extraction and a subsequent branch module for training various recognition functions.

[0050] In some embodiments, the gait model includes multi-layer convolutional modules for sequence feature extraction, multiple branch modules for gait sequence number recognition training and gait angle recognition training, and function training set according to actual conditions; for example, three convolutional modules are set for sequence feature extraction, each convolutional module consists of two layers of 3D convolution, followed by a first branch and a second branch, which are used for gait sequence number recognition training and gait angle recognition training, respectively, to determine a first classification feature vector and a first angle feature vector; wherein, the first classification feature vector includes temporal features and static features, the temporal features are extracted based on the lower body torso, and the static features are extracted based on the upper body torso.

[0051] Alternatively, the recognition function training can also be scene recognition function training or pitch angle recognition function training, etc., and can be set accordingly according to the actual situation.

[0052] S22. Obtain the second classification feature vector and the second angle feature vector of each candidate gait sequence from at least two candidate gait sequences;

[0053] The gait library contains several candidate gait sequences. The gait library is formed based on several candidate gait sequences. The second classification feature vector and the second angle feature vector of each candidate gait sequence are obtained, just like the acquisition step in step S21.

[0054] In some embodiments, the pre-training of the gait model is as follows: acquiring full-view gait data of the target site and creating label data based on the full-view gait data; determining the contour sequence based on the full-view gait data and label data; inputting the contour sequence and extracting the sequence features of the contour sequence through a convolution module; extracting the temporal and static features of the contour sequence through the first branch of the gait model based on the sequence features; extracting the angular features of the contour sequence through the second branch of the gait model based on the sequence features; training the gait sequence number recognition function of the model to be trained according to the temporal and static features; training the gait angle recognition function of the model to be trained according to the angular features; wherein, the contour sequence corresponds to the gait sequence, that is, the gait sequence is determined by the contour of a pedestrian walking; by synchronously acquiring the temporal and static features of the relevant sequences, the cross-view gait recognition effect is improved, the classification ability of gait features is enhanced, and the recognition accuracy is improved.

[0055] In some embodiments, full-view gait data of the target site is acquired, and label data is generated based on the full-view gait data. A contour sequence is determined based on the full-view gait data and the label data. The label data includes gait angle labels and gait sequence number labels.

[0056] Among them, the target site is the site where the full-view gait data is acquired; the full-view gait data is the gait data acquired from a 360-degree view; the gait angle label is the angle label related to the camera device; and the gait sequence number label is the pedestrian ID.

[0057] Specifically, a camera system is set up at the target site to acquire gait data from all angles. Gait angle labels and gait sequence labels are assigned to each angle and each pedestrian. Then, the contour sequences corresponding to the gait data from all angles are determined based on these data. For example, a ring of cameras is set up around the target site, and different pedestrians walk back and forth within the site, collecting corresponding gait data from all angles. This allows for the determination of several contour sequences corresponding to the gait data from all angles, thus acquiring sufficient viewpoint data to improve the accuracy of cross-angle recognition.

[0058] In some embodiments, the label data may also include height labels, scene labels, pitch angle labels, etc., which can be set according to actual needs;

[0059] In some embodiments, the gait angle label is generated according to the angular relationship between the person's walking direction and the camera. The scene label can be generated based on the camera's installation height, and the pitch angle label can be generated based on the angle between the camera's optical axis and the horizon. The scene label can be for daytime imaging, nighttime color imaging, nighttime infrared imaging, etc., in the same scene, or for different shooting scenes, such as indoor shopping malls, outdoor grass, outdoor concrete, outdoor uneven ground, etc. For example, the target site is a circle, and pedestrians walk back and forth in the circle. A ring of cameras is set around the circle, and the cameras all shoot towards the center of the circle to obtain relevant data. Each camera has a corresponding height label, angle label, and pitch angle label.

[0060] For example: the contour sequence is input into the network of the model to be trained, and sequence features are extracted through three convolutional models to obtain the sequence features corresponding to the contour sequence; the first branch of the model to be trained extracts the corresponding temporal features and static features based on the sequence features, and its output feature dimension is (b,c2,h2,w2), and then global average pooling is used to transform the features into (b,c2) to generate a second classification feature vector F1 with a length of c2; the second branch of the model to be trained extracts the corresponding angular features based on the sequence features, and uses max pooling to fuse multiple frames into a single frame with a feature dimension of (b,c2,h2,w2), and then uses global average pooling to transform the features into (b,c2) to generate a second angular feature vector F2 with a length of c2.

[0061] Global average pooling is the process of averaging all pixel values ​​in each channel of the feature map to obtain a new 1*1 channel map. For example, if the input is (N,C,H,W), the output is (N,C,1,1). Max pooling divides the input image into several rectangular regions and outputs the maximum value for each sub-region.

[0062] Optionally, the gait sequence recognition function of the model to be trained can be trained using cross-entropy ID classification loss and triplet loss, or the appropriate loss training method can be selected according to actual needs.

[0063] Optionally, the gait angle recognition function of the model to be trained can be trained by using cross-entropy loss to train viewpoint classification, or by training scene classification or pitch angle classification according to actual needs.

[0064] In some embodiments, the first branch of the model to be trained includes a feature extraction and fusion module, which comprises a first feature extraction module and a second feature extraction module. The feature map in the contour sequence is segmented into an upper torso portion and a lower torso portion. The first feature extraction module is assigned to the upper torso portion of the feature map, and the second feature extraction module is assigned to the lower torso portion. The first feature extraction module uses a 3D convolutional layer combined with a max-pooling layer to extract static features, and its feature dimension is adjusted from (b, c). The static features and temporal features are then concatenated and fed into the subsequent global average pooling module, which transforms the feature dimensions from (b,c2,t,h3,w2) to (b,c2,h3,w2).

[0065] S3. Using the first classification feature vector and the second classification feature vector of each candidate gait sequence, calculate the first distance matrix feature corresponding to each candidate gait sequence; and using the first angle feature vector and the second angle feature vector of each candidate gait sequence, calculate the second distance matrix feature corresponding to each candidate gait sequence.

[0066] Wherein, the first classification feature vector is the feature vector corresponding to the gait sequence to be detected, and its length is C2; the second classification feature vector is the feature vector corresponding to the candidate gait sequence in the gait library, and its length is C2; the first angle feature vector is the feature vector corresponding to the viewpoint classification of the gait sequence to be detected, and its length is C2; the second angle feature vector is the feature vector corresponding to the viewpoint classification of the candidate gait sequence in the gait library, and its length is C2.

[0067] See Figure 4 , Figure 4 This is a flowchart illustrating one embodiment of step S3, which includes:

[0068] S31. Obtain the distance between the first classification feature vector and the second classification feature vector of each candidate gait sequence;

[0069] Wherein, the first classification feature vector is the feature vector corresponding to the gait sequence to be detected, and the second classification feature vector is the feature vector corresponding to at least two candidate gait sequences in the gait library. The distance between them is the distance between the feature vector corresponding to the gait sequence to be detected and the feature vector corresponding to each candidate gait sequence in the gait library.

[0070] In some embodiments, the distance is calculated by measuring the distance between the first classification feature vector and the second classification feature vector corresponding to all candidate gait sequences in the gait database.

[0071] Optionally, the distance calculation method can be the cosine distance calculation method, or a distance calculation method selected according to the actual situation.

[0072] S32. Calculate the first distance matrix features corresponding to each candidate gait sequence using the obtained distances;

[0073] The distance obtained is the distance between the first classification feature vector and the second classification feature vector corresponding to each candidate gait sequence in the gait library. Therefore, it is a series of distances, which in turn form a distance matrix.

[0074] In some embodiments, a series of distances between the first classification feature vector obtained by the cosine distance calculation method and the second classification feature vector corresponding to each candidate gait sequence in the gait library are used to form the first distance matrix feature corresponding to each candidate gait sequence.

[0075] In some embodiments, cosine distance can also be called cosine similarity. The cosine value ranges from [-1, 1]. The angle between two vectors is calculated, and the corresponding cosine value is obtained. This cosine value can be used to represent the similarity between the two vectors. The smaller the angle, approaching 0 degrees, the closer the cosine value is to 1, and the more similar their directions are. When the directions of the two vectors are completely opposite, the cosine of the angle reaches its minimum value of -1. When the cosine value is 0, the two vectors are orthogonal, and the angle between them is 90 degrees. Therefore, it can be seen that cosine similarity is independent of the magnitude of the vectors and only related to their direction. In this embodiment, the cosine distance range is set to [0:2]. If the perspectives are the same, the scene is the same, or the pitch angle is the same, the cosine distance is smaller; if the perspectives are different, the cosine distance is larger.

[0076] See Figure 5 , Figure 5 This is a flowchart illustrating another embodiment of step S3, including:

[0077] S33. Obtain the distance between the first angle feature vector and the second angle feature vector of each candidate gait sequence;

[0078] Wherein, the first angle feature vector is the feature vector corresponding to the viewpoint classification of the gait sequence to be detected, the second angle feature vector is the feature vector corresponding to the viewpoint classification of each candidate gait sequence in the gait library, and the distance is the distance between the feature vector corresponding to the viewpoint classification of the gait sequence to be detected and the feature vector corresponding to the viewpoint classification of each candidate gait sequence in the gait library.

[0079] In some embodiments, the distance is calculated by measuring the distance between the first angular feature vector and the second angular feature vector corresponding to each candidate gait sequence in the gait library.

[0080] Optionally, the distance calculation method can be the cosine distance calculation method, or a distance calculation method selected according to the actual situation.

[0081] S34. Using the obtained distances, calculate the second distance matrix features corresponding to each candidate gait sequence;

[0082] The distance obtained is the distance between the first angular feature vector and the second angular feature vector corresponding to each candidate gait sequence in the gait library. Therefore, it is a series of distances, which in turn form a distance matrix.

[0083] In some embodiments, a series of distances between the first angular feature vector obtained by the cosine distance calculation method and the second angular feature vector corresponding to each candidate gait sequence in the gait library are used to form the second distance matrix feature corresponding to each candidate gait sequence.

[0084] S4. Using preset fusion weights, fuse the features of the first distance matrix and the features of the second distance matrix to obtain the feature distance of the fused angle information corresponding to each candidate gait sequence;

[0085] The fusion process involves using fusion weights to calculate the features of the first and second distance matrices and obtain the calculation results. Since the features of the second distance matrix correspond to a series of distances between the first angular feature vector and the second angular feature vectors corresponding to each candidate gait sequence in the gait library, the fusion calculation result is the feature distance of the fused angular information.

[0086] In some embodiments, the first distance matrix features and the second distance matrix features with preset fusion weights are used to perform calculation operations to obtain the feature distance of the fusion angle information corresponding to each candidate gait sequence.

[0087] See Figure 6 , Figure 6 This is a flowchart illustrating one embodiment of step S4, which includes:

[0088] S41. Set preset fusion weights for the features of the second distance matrix;

[0089] In this context, since the second distance matrix features correspond to a series of distances between the first angle feature vector and the second angle feature vector corresponding to each candidate gait sequence in the gait library, in order to obtain the feature distance of the fused angle information, it is necessary to set a preset fusion weight for the second angle feature vector so that the distinction between sequences with the same viewpoint and sequences with different viewpoints can be amplified.

[0090] In some embodiments, the similarity that needs to be subtracted from the same viewpoint is smaller, while the similarity that needs to be subtracted from different viewpoints is larger.

[0091] S42. The first distance matrix feature and the second distance matrix feature with preset fusion weights are calculated to obtain the feature distance of the fused angle information. The feature distance of the same state angle label is less than the feature distance of different state angle labels.

[0092] The calculation operation can be a scoring operation to facilitate cross-view recognition. This is because the same gait number has low similarity under different viewpoints, scenes, or pitch angles, while different gait numbers may have slightly higher similarity under the same viewpoint, scene, or pitch angle. This can easily lead to gait recognition failures when crossing viewpoints. Therefore, it is necessary to score the similarity of samples from different viewpoints to facilitate cross-view recognition.

[0093] Optionally, if similarity is scored by adding points, it may exceed the maximum score, thus affecting the recognition result. Therefore, the scoring operation is converted into a subtraction operation, that is, the similarity score is reduced by a little more when the same viewpoint is used, and the similarity score is reduced by a little less when the viewpoint is used.

[0094] In some embodiments, the first distance matrix feature is set as D1, the second distance matrix feature is set as D2, and the preset fusion weight is set to λ, where λ ranges from 0 to 1. The feature distance of the fused angle information is set to D3. Then, we have: D1 - λ * D2 = D3. According to this formula, the distance subtraction for the same viewpoint is smaller, while the distance subtraction for different viewpoints is larger. The smaller the distance, the higher the similarity, which is equivalent to adding points to different viewpoints.

[0095] In some embodiments, the fusion weight is used to address situations where inter-class similarity may be greater than intra-class similarity. That is, contour sequences are significantly affected by viewing angle, and inter-class similarity may be greater than intra-class similarity. For example, the same gait number may have low similarity under different viewing angles, scenes, or pitch angles, while different gait numbers may have slightly higher similarity under the same viewing angle, scene, or pitch angle. This can easily lead to gait recognition failures when performing cross-view recognition. Therefore, it is necessary to add weights to the similarity scores of samples from different viewing angles to facilitate cross-view recognition.

[0096] Optionally, if similarity is scored by adding points, it may exceed the maximum score, thus affecting the recognition result. Therefore, the scoring operation is converted into a subtraction operation, that is, the similarity score is reduced by a little more when the same viewpoint is used, and the similarity score is reduced by a little less when the viewpoint is used.

[0097] In some embodiments, based on the features of the second distance matrix, the fusion weight of synchronous angle labels is set to be greater than the fusion weight of asynchronous angle labels; for example, the weight value is set to λ, and the value of the weight value λ is between 0 and 1, such as 0, 0.1, 0.2, 0.5, 0.7, 0.8, 0.9, 1, etc., and the corresponding fusion weight is set according to the features of the second distance matrix.

[0098] In some embodiments, the preset fusion weight is a fixed value, ranging from 0 to 1, with a preferred value of 0.1. The second distance matrix automatically calculates the distance between labels with the same angle and labels with different angles. Furthermore, in the distance matrix: the distance value for labels with the same angle is small, and the distance value for labels with different angles is large. The angles are also automatically output by the pre-trained gait model.

[0099] S5. Reorder at least two candidate gait sequences according to the feature distances between each candidate gait sequence and the gait sequence to be detected.

[0100] The gait database contains several candidate gait sequences. Each candidate gait sequence has a corresponding feature distance to the gait sequence to be detected. Therefore, a series of feature distances can be obtained. Based on this series of feature distances, the candidate gait sequences can be reordered.

[0101] Specifically, the feature distances between each candidate gait sequence in the gait database and the gait sequence to be detected are obtained. The obtained series of feature distances are sorted by size, and the candidate gait sequences in the gait database are reordered based on the sorting results of this series of feature distances.

[0102] See Figure 7 , Figure 7 This is a flowchart illustrating an embodiment of step S5, which includes:

[0103] S51. Obtain the feature distance of the fusion angle information of each candidate gait sequence in at least two candidate gait sequences;

[0104] Among them, at least two candidate gait sequences are candidate gait sequences in the gait library.

[0105] Specifically, if there are two candidate gait sequences in the gait library, then the feature distances of the two fused angle information are obtained, and these two feature distances correspond to the two candidate gait sequences; if there are several candidate gait sequences in the gait library, then several feature distances of the fused angle information need to be obtained, and these several feature distances correspond to several candidate gait sequences.

[0106] S52. Based on the feature distances between each candidate gait sequence and the gait sequence to be detected, reorder at least two candidate gait sequences to determine the reordered candidate gait sequences.

[0107] The feature distance between synchronous angle labels is smaller than that between asynchronous angle labels. This is to ensure that candidate gait sequences are closer when the angle labels are synchronous.

[0108] Specifically, a series of feature distances are obtained between each candidate gait sequence and the gait sequence to be detected. The obtained series of feature distances are sorted by size. The candidate gait sequences in the gait library are reordered based on the sorting results of this series of feature distances, and then the reordered candidate gait sequences are determined.

[0109] S6. Perform gait recognition on the target based on the reordered candidate gait sequence and the gait sequence to be detected.

[0110] The gait sequence to be detected is the gait sequence corresponding to the target. Therefore, by matching the reordered candidate gait sequence with the gait sequence to be detected, the gait recognition result of the target can be obtained.

[0111] In some embodiments, a reordered candidate gait sequence is obtained; a feature distance comparison is performed on the gait sequence to be detected based on the reordered candidate gait sequence; and a gait recognition result is determined based on the feature distance comparison result.

[0112] In some embodiments, the feature distances corresponding to several candidate gait sequences in the gait database and the gait sequence to be detected are sorted by size. The feature distances corresponding to each candidate gait sequence are compared, and the smallest feature distance is obtained to obtain the recognition result of the gait sequence to be detected. That is, the smaller the feature distance, the higher the probability that they belong to the same sequence label and the higher their similarity.

[0113] For example: The gait sequence to be detected is sample A. The gait database contains candidate walking sequences samples B and C. Samples B and A have the same viewpoint but different sequence numbers. The first distance matrix feature corresponding to sample B is 0.8. Samples C and A have different viewpoints but the same sequence number. The first distance matrix feature corresponding to sample C is 0.9. If only the first distance matrix feature is used for identification, i.e., without angle information, sample B, with the same viewpoint but different sequence number, will be ranked before sample C, leading to an error in gait recognition. However, if angle information is combined, the second distance matrix feature corresponding to sample B, with a preset fusion weight, is 0.1, and the second distance matrix feature corresponding to sample C, with a preset fusion weight, is 0.3. Therefore, the feature distance between sample B and sample A is 0.8 - 0.1 = 0.7, and the feature distance between sample C and sample A is 0.9 - 0.3 = 0.6. Thus, sample C is ranked before sample B, and the gait recognition is correct. That is, sample C is the closest to sample A, and the candidate gait sequence corresponding to sample C is the closest to the gait sequence to be detected.

[0114] Unlike existing technologies, this embodiment acquires the gait sequence to be detected and at least two candidate gait sequences from a gait database; extracts a first classification feature vector and a first angle feature vector from the gait sequence to be detected, and extracts a second classification feature vector and a second angle feature vector from the at least two candidate gait sequences; calculates a first distance matrix feature corresponding to each candidate gait sequence using the first classification feature vector and the second classification feature vector of each candidate gait sequence; calculates a second distance matrix feature corresponding to each candidate gait sequence using the first angle feature vector and the second angle feature vector of each candidate gait sequence; fuses the first distance matrix feature and the second distance matrix feature using a preset fusion weight to obtain the feature distance of the fused angle information corresponding to each candidate gait sequence; and matches each candidate gait sequence with the gait sequence to be detected. The feature distance is used to reorder at least two candidate gait sequences; the gait recognition of the target is performed based on the reordered candidate gait sequences and the gait sequence to be detected; that is, this application extracts the first classification feature vector and the first angle feature vector of the gait sequence to be detected, as well as the second classification feature vector and the second angle feature vector corresponding to each candidate gait sequence in at least two candidate gait sequences in the gait library, and then obtains the first distance matrix feature and the second distance matrix feature corresponding to each candidate gait sequence; and obtains the feature distance of the fusion angle corresponding to each candidate gait sequence based on the first distance matrix feature and the second distance matrix feature, and reorders the viewpoints according to the feature distances corresponding to each candidate gait sequence and the gait sequence to be detected, and completes the recognition of the gait sequence to be detected based on the reordered candidate gait sequences. All viewpoints can be reordered at the same time, which improves the recognition efficiency.

[0115] Please see Figure 8 , Figure 8 This is a schematic diagram of an embodiment of the gait recognition system of the present invention. This system can perform the steps of the gait recognition method described above. For related details, please refer to the detailed description in the above method; it will not be repeated here.

[0116] The gait recognition system 200 includes: an acquisition module 210, an extraction module 220, a first calculation module 230, a second calculation module 240, a fusion module 250, a re-ranking module 260, and a recognition module 270; wherein, the acquisition module 210 is used to acquire the gait sequence to be detected and to acquire at least two candidate gait sequences from a gait database; the extraction module 220 uses a pre-trained gait model to extract a first classification feature vector and a first angle feature vector of the gait sequence to be detected, and to extract a second classification feature vector and a second angle feature vector of the at least two candidate gait sequences; the first calculation module 230 uses the first classification feature vector and the second classification feature vector of each candidate gait sequence... The first distance matrix feature corresponding to each candidate gait sequence is calculated; the second calculation module 240 uses the first angle feature vector and the second angle feature vector of each candidate gait sequence to calculate the second distance matrix feature corresponding to each candidate gait sequence; the fusion module 250 uses fusion weights to fuse the first distance matrix feature and the second distance matrix feature to obtain the feature distance of the fused angle information corresponding to each candidate gait sequence; the reordering module 260 reorders at least two candidate gait sequences according to the feature distances corresponding to each candidate gait sequence and the gait sequence to be detected; the recognition module 270 performs gait recognition on the target based on the reordered candidate gait sequences and the gait sequence to be detected.

[0117] In some embodiments, the gait recognition system further includes a gait model, which is pre-trained; the gait model includes at least one convolutional module, and a first branch module and a second branch module respectively connected to the at least one convolutional module; the first branch module includes a feature extraction and fusion module, a first pooling module, a first loss training module, and a second loss training module connected in sequence, and the first branch module is used to train the gait sequence number recognition function of the gait sequence to be detected; wherein, the feature extraction and fusion module is used to extract the temporal features and static features of the gait sequence to be detected; the second branch module includes a second pooling module, a third pooling module, and a third loss training module connected in sequence, and the second branch module is used to train the gait angle recognition function of the gait sequence to be detected.

[0118] In some embodiments, the feature extraction and fusion module includes: a first feature extraction module and a second feature extraction module; the first feature extraction module is used to extract static features, and the second feature extraction module is used to extract temporal features; wherein, the temporal features are extracted based on the lower body torso, and the static features are extracted based on the upper body torso.

[0119] Please see Figure 9 , Figure 9 This is a schematic diagram of an embodiment of the electronic device of the present invention. This electronic device can perform the steps in the above-described gait recognition method.

[0120] The electronic device 300 includes a memory 320 and a processor 310 coupled to the memory. The memory 320 stores program instructions. The processor 310 executes the program instructions to implement the steps in the gait recognition method described above. For related details, please refer to the detailed description of the gait recognition method described above, which will not be repeated here.

[0121] Please see Figure 10 , Figure 10 This is a schematic diagram of a computer-readable storage medium according to an embodiment of the present invention. The computer-readable storage medium 400 stores a computer program 410, which, when executed by a processor, implements the steps in the gait recognition method described above. For related details, please refer to the detailed description of the method described above; it will not be repeated here.

[0122] The above scheme obtains the first classification feature vector and the first angle feature vector of the gait sequence to be detected, as well as the second classification feature vector and the second angle feature vector corresponding to each candidate gait sequence in at least two candidate gait sequences in the gait database. Then, it obtains the first distance matrix feature and the second distance matrix feature corresponding to each candidate gait sequence. Based on the first distance matrix feature and the second distance matrix feature, it obtains the feature distance of the fusion angle corresponding to each candidate gait sequence. The viewpoints are then reordered according to the feature distances between each candidate gait sequence and the gait sequence to be detected. The recognition of the gait sequence to be detected is completed based on the reordered candidate gait sequences. This scheme can simultaneously reorder all viewpoints, thereby improving recognition efficiency.

[0123] Furthermore, it can simultaneously acquire temporal and static features of relevant sequences, thereby improving cross-view gait recognition performance, enhancing the classification ability of gait features, and improving recognition accuracy.

[0124] In the several embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, indirect coupling or communication connection of apparatuses or units, and may be electrical, mechanical, or other forms.

[0125] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0126] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0127] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0128] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A gait recognition method characterized by comprising: The gait recognition method comprises: obtaining a to-be-detected gait sequence and at least two candidate gait sequences in a gait library; extracting a first classification feature vector and a first angle feature vector of the to-be-detected gait sequence, and extracting a second classification feature vector and a second angle feature vector of each candidate gait sequence in the at least two candidate gait sequences, wherein the first classification feature vector is obtained by converting time sequence features and static features extracted from the to-be-detected gait sequence, and the first angle feature vector is obtained by converting angle features extracted from the to-be-detected gait sequence; obtaining a first distance matrix feature corresponding to each candidate gait sequence by using the first classification feature vector and the second classification feature vector of each candidate gait sequence, and calculating a second distance matrix feature corresponding to each candidate gait sequence by using the first angle feature vector and the second angle feature vector of each candidate gait sequence; fusing the first distance matrix feature and the second distance matrix feature by using a preset fusion weight to obtain a feature distance of fusion angle information corresponding to each candidate gait sequence; reordering the at least two candidate gait sequences according to the feature distances corresponding to the to-be-detected gait sequence and the respective candidate gait sequences; performing gait recognition on a target according to the to-be-detected gait sequence and the reordered candidate gait sequences.

2. The gait recognition method of claim 1, wherein: the obtaining of the to-be-detected gait sequence and the at least two candidate gait sequences in the gait library comprises: obtaining pedestrian gait data to be detected, and determining the to-be-detected gait sequence based on the pedestrian gait data; obtaining various types of gait data in real scenes, determining a plurality of candidate gait sequences based on the various types of gait data in real scenes, and then composing a gait library; the extracting of the first classification feature vector and the first angle feature vector of the to-be-detected gait sequence, and the extracting of the second classification feature vector and the second angle feature vector of each candidate gait sequence in the at least two candidate gait sequences comprises: extracting the first classification feature vector and the first angle feature vector of the to-be-detected gait sequence, and extracting the second classification feature vector and the second angle feature vector of each candidate gait sequence in the at least two candidate gait sequences by using a pre-trained gait model.

3. The gait recognition method of claim 2, wherein: the extracting of the first classification feature vector and the first angle feature vector of the to-be-detected gait sequence by using the pre-trained gait model comprises: extracting a first classification feature vector of the to-be-detected gait sequence by using a first branch of the pre-trained gait model, and extracting a first angle feature vector of the to-be-detected gait sequence by using a second branch of the pre-trained gait model; wherein the first classification feature vector comprises time sequence features and static features, the time sequence features are extracted based on a lower body trunk, and the static features are extracted based on an upper body trunk. The second classification feature vector and the second angle feature vector of each of the at least two candidate gait sequences are extracted. The second classification feature vector and the second angle feature vector are determined based on each of the at least two candidate gait sequences in the gait library.

4. The gait recognition method of claim 1, wherein, The first distance matrix feature corresponding to each candidate gait sequence is calculated by using the first classification feature vector and the second classification feature vector of each candidate gait sequence, and includes: The distance between the first classification feature vector and the second classification feature vector of each candidate gait sequence is obtained, and the first distance matrix feature corresponding to each candidate gait sequence is obtained based on the obtained distance; The second distance matrix feature corresponding to each candidate gait sequence is calculated by using the first angle feature vector and the second angle feature vector of each candidate gait sequence, and includes: The distance between the first angle feature vector and the second angle feature vector of each candidate gait sequence is obtained, and the second distance matrix feature corresponding to each candidate gait sequence is obtained based on the obtained distance.

5. The gait recognition method of claim 1, wherein, Before the first distance matrix feature and the second distance matrix feature are fused by using the preset fusion weight, the gait recognition method further includes: Based on the second distance matrix feature, the preset fusion weight is set to have a value range of 0-1.

6. The gait recognition method of claim 1, wherein, The first distance matrix feature and the second distance matrix feature are fused by using the preset fusion weight to obtain the feature distance of the fusion angle information corresponding to each candidate gait sequence, and includes: The first distance matrix feature and the second distance matrix feature with the preset fusion weight are calculated to obtain the feature distance of the fusion angle information corresponding to each candidate gait sequence.

7. The gait recognition method of claim 1, wherein, The at least two candidate gait sequences are reordered according to the feature distance corresponding to each candidate gait sequence and the to-be-detected gait sequence, and includes: The feature distance of the fusion angle information of each candidate gait sequence in the at least two candidate gait sequences is obtained; The at least two candidate gait sequences are size-reordered according to the feature distance corresponding to each candidate gait sequence and the to-be-detected gait sequence to determine the reordered candidate gait sequence; wherein the feature distance of the same gait angle label is smaller than the feature distance of different gait angle labels.

8. The gait recognition method of claim 7, wherein, The gait recognition is performed on the to-be-detected gait sequence according to the reordered candidate gait sequence and the to-be-detected gait sequence, and includes: The reordered candidate gait sequence is obtained; The to-be-detected gait sequence is compared in feature distance according to the reordered candidate gait sequence, and the gait recognition result is determined based on the comparison result in feature distance.

9. The gait recognition method of claim 1, wherein, The pre-trained gait model comprises: Obtaining full-view gait data of a target site, and making label data according to the full-view gait data, determining a contour sequence based on the full-view gait data and the label data; wherein the label data comprises a gait angle label and a gait sequence number label; Inputting the contour sequence, and extracting sequence features of the contour sequence; Based on the sequence features, extracting time sequence features and static features of the contour sequence; Based on the sequence features, extracting angle features of the contour sequence; According to the time sequence features and the static features, training a gait sequence number recognition function of a to-be-trained model; According to the angle features, training a gait angle recognition function of the to-be-trained model.

10. The gait recognition method according to claim 9, characterized in that: The training of the gait sequence number recognition function of the to-be-trained model comprises triad loss training and gait sequence number classification loss training of the gait sequence number recognition function of the to-be-trained model; The training of the gait angle recognition function of the to-be-trained model comprises training of a viewing angle or a scene or a pitch angle of the to-be-trained model.

11. A gait recognition system characterized by, Comprise: An acquisition module is configured to acquire a to-be-detected gait sequence and at least two candidate gait sequences in a gait library; An extraction module is configured to extract a first classification feature vector and a first angle feature vector of the to-be-detected gait sequence, and extract a second classification feature vector and a second angle feature vector of each candidate gait sequence in the at least two candidate gait sequences, wherein the first classification feature vector is obtained by converting time sequence features and static features extracted from the to-be-detected gait sequence, and the first angle feature vector is obtained by converting angle features extracted from the to-be-detected gait sequence; A first calculation module is configured to calculate a first distance matrix feature corresponding to each candidate gait sequence by using the first classification feature vector and the second classification feature vector of each candidate gait sequence; A second calculation module is configured to calculate a second distance matrix feature corresponding to each candidate gait sequence by using the first angle feature vector and the second angle feature vector of each candidate gait sequence; A fusion module is configured to fuse the first distance matrix feature and the second distance matrix feature by using a preset fusion weight to obtain a feature distance of fusion angle information corresponding to each candidate gait sequence; A reordering module is configured to reorder the at least two candidate gait sequences according to the feature distances corresponding to the at least two candidate gait sequences and the to-be-detected gait sequence respectively; An identification module is configured to perform gait recognition on a target according to the reordered candidate gait sequences and the to-be-detected gait sequence.

12. The gait recognition system of claim 11, wherein, Further comprise: A gait model, wherein the gait model is pre-trained; The gait model comprises at least one convolution module, and a first branch module and a second branch module connected with the at least one convolution module respectively; The first branch module comprises a feature extraction and fusion module, a first pooling module, a first loss training module and a second loss training module connected in sequence, and is configured to perform gait number identification function training on the to-be-detected gait sequence; the feature extraction and fusion module is configured to extract time sequence features and static features of the to-be-detected gait sequence. The second branch module comprises a second pooling module, a third pooling module and a third loss training module connected in sequence, and is configured to perform gait angle identification function training on the to-be-detected gait sequence.

13. The gait recognition system of claim 12, wherein, The feature extraction and fusion module comprises a first feature extraction module and a second feature extraction module; the first feature extraction module is configured to extract static features, and the second feature extraction module is configured to extract time sequence features; the time sequence features are extracted based on a lower body trunk, and the static features are extracted based on an upper body trunk.

14. An electronic device, comprising: A computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps of the gait recognition method according to any one of claims 1 to 10.

15. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps of the gait recognition method according to any one of claims 1 to 10.

Images