Gait recognition network training method and system, and gait recognition method

Abstract

The application relates to a gait recognition network training method and system and a gait recognition method. The gait recognition network training method comprises the following steps: obtaining M groups of training data, one group of training data comprising two gait energy maps; horizontally dividing the gait energy maps into N parts; simultaneously extracting block features of the N parts of the gait energy maps through a basic convolution layer; sending the block features into a second convolution layer for processing; constructing a loss function based on the output results of the convolution layers, and supervising and training the gait recognition network through the loss function. Each image is horizontally divided and subjected to block convolution, so that even if the legs are blocked, the features of other parts are still distinguishable, and therefore the pedestrian gait recognition accuracy can be improved; the two images in each group are spliced, and block convolution is further performed on the spliced features, so that the loss function can make the features more discriminative, thereby improving the recognition accuracy of the model.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, specifically to a training method, system, and gait recognition method for a gait recognition network. Background Technology

[0002] Pedestrian gait recognition verifies a pedestrian based on the behavioral characteristics exhibited by various parts of their body, especially their legs, while walking. Traditional pedestrian gait recognition methods primarily rely on a combination of manually extracted features, such as facial features and optical flow characteristics generated during movement, from video frames, and employ classifiers like SVM (Support Vector Machine) to classify and discriminate these features. With the increasing maturity of convolutional neural networks (CNNs) in image processing, CNN-based pedestrian gait recognition methods have emerged. These methods mainly involve various image input methods: including RGB image sequences from the original video, binary image sequences after pedestrian segmentation, and contour energy map sequences combining binary image sequences. Different CNN structures are designed for different image types to extract pedestrian gait features, thereby significantly improving the accuracy and recognition speed of gait recognition methods.

[0003] Current pedestrian gait recognition methods mainly involve three steps: pedestrian detection, pedestrian segmentation, and pedestrian gait recognition in video frames or images. The goal is to determine whether pedestrians captured by cameras in different scenes or across different areas are the same person. Due to the different types of input images, the specific implementation of pedestrian gait recognition methods also varies.

[0004] In related technologies, the methods mentioned above use pedestrian binary contour maps as input. However, in real-world scenarios, pedestrian postures vary greatly. When pedestrians carry backpacks, wear long coats, or carry objects, it directly interferes with the pedestrian's contour features, leading to inaccurate gait recognition and affecting pedestrian verification. Furthermore, a single binary map does not contain the temporal features of pedestrian movement and is insufficient to distinguish gait changes between different people. Summary of the Invention

[0005] To overcome, to some extent, the problems in related technologies, such as the susceptibility of pedestrian contour features to interference leading to inaccurate gait recognition and the lack of temporal features of pedestrian movement in a single binary image, this application provides a training method, system, and gait recognition method for a gait recognition network.

[0006] According to a first aspect of the embodiments of this application, a method for training a gait recognition network is provided, comprising:

[0007] Obtain M sets of training data, where each set of training data includes two gait energy maps, and M≥3;

[0008] Divide the vertical height of the gait energy map into N parts, where N≥4;

[0009] The N parts of the gait energy map are simultaneously passed through a basic convolutional layer to extract block features;

[0010] The block features are fed into the second convolutional layer for processing. The second convolutional layer includes two branches: the first branch performs convolution on the block features of each image, and the second branch performs convolution on the concatenated block features of two images in a set.

[0011] A loss function is constructed based on the output of each convolutional layer, and the gait recognition network is trained under supervision using the loss function.

[0012] Furthermore, the steps for obtaining training data include:

[0013] Video streams are acquired from the monitoring system, and pedestrian detection and segmentation are performed based on video frames to obtain a binary contour sequence map containing only pedestrians.

[0014] Different pedestrians are numbered, with each pedestrian corresponding to a different ID number, and each ID number corresponding to multiple binary contour sequence images;

[0015] Generate the corresponding gait energy map based on the binary contour sequence map.

[0016] Furthermore, if M is set to three, then obtaining M sets of training data involves the following steps:

[0017] Four different ID numbers are randomly selected and labeled as id1, id2, id3 and id4;

[0018] Select three gait energy maps from id1, and select one gait energy map from each of the other three ID numbers;

[0019] Use the two gait energy maps of id1 as the first set of training data;

[0020] The remaining gait energy map of id1 and the gait energy maps of id2, id3, and id4 are paired and used as the second and third sets of training data.

[0021] Furthermore, a loss function is constructed based on the output of each convolutional layer, including:

[0022] Construct a five-fold loss function, which is the sum of two terms;

[0023] The first term is: the triplet loss function L among the three sets of training data. tripThe combined loss is: the distance between the concatenated features of the first set of training data and the concatenated features of the second set of training data, minus the distance between the concatenated features of the second set of training data and the concatenated features of the third set of training data.

[0024] Furthermore, the triplet loss function L trip for:

[0025] L trip (A,P,N)=max(d(A,P)-d(A,N)+α,0);

[0026] Where α represents the minimum feature interval distance; This indicates that the distance between x and y is calculated.

[0027] Furthermore, the image feature loss function L itself imf for:

[0028] L imf =η1L trip (x id1 ,x' id1 ,x id2 )+η2L trip (x id1 ,x' id1 ,x id3 )+η3L trip (x id1 ,x' id1 ,x id4 );

[0029] η1, η2, and η3 are configurable hyperparameters, representing the weight of the expression multiplied by them in the overall loss.

[0030] Furthermore, the image feature stitching feature loss function L conc for:

[0031]

[0032] This represents the image stitching features of the first set of training data. These are the image stitching features from the second set of training data. These are the image stitching features of the third set of training data.

[0033] Furthermore, the five-fold loss function L is:

[0034] L = η imf L imf +η conc L conc ;

[0035] Where η imfIt is the image feature loss function L imf The weighting parameter, η conc It is the concatenation feature loss function L conc The weight parameters.

[0036] According to a second aspect of the embodiments of this application, a training system for a gait recognition network is provided, comprising:

[0037] The acquisition module is used to acquire M sets of training data. Each set of training data includes two gait energy maps, where M≥3.

[0038] The segmentation module is used to divide the vertical height of the gait energy map into N parts, where N≥4;

[0039] The basic convolutional module is used to extract block features by simultaneously passing N parts of the gait energy map through the basic convolutional layer;

[0040] The second convolution module is used to feed the block features into the second convolution layer for processing. The second convolution layer includes two branches: the first branch performs convolution on the block features of each image, and the second branch performs convolution on the concatenated block features of two images in a set.

[0041] The supervised training module is used to construct a loss function based on the output of each convolutional layer, and to supervise the training of the gait recognition network through the loss function.

[0042] According to a third aspect of the embodiments of this application, a gait recognition method is provided, comprising:

[0043] Retrieve camera video, perform pedestrian detection on video frames, and establish a pedestrian ID database;

[0044] For each video sequence with a given ID, target segmentation is performed to separate pedestrians and background.

[0045] During target segmentation, a sequence of binary contour maps of pedestrians is output, and the GEI map of the pedestrians is calculated based on the binary maps.

[0046] The GEI image of a pedestrian is input into a gait recognition network to extract gait features and establish a basic database of pedestrian gait features.

[0047] When a pedestrian video frame to be queried is input, gait features are first extracted through a gait recognition network, and then similarity is measured with the features in the pedestrian gait feature database to find all video frames of the same person.

[0048] The gait recognition network is obtained through the training method of the gait recognition network described in any of the above embodiments.

[0049] The technical solutions provided by the embodiments of this application have the following beneficial effects:

[0050] This application's scheme uses GEI images as network input, performs block convolution on each image after horizontal segmentation, and outputs the concatenated features. When a pedestrian moves, their leg state changes significantly, but when the legs are occluded, the features cannot represent the pedestrian's gait. Therefore, this scheme divides the image horizontally into multiple blocks. Even if the legs are occluded, the features of other parts remain distinguishable, thus improving the accuracy of pedestrian gait recognition. Furthermore, due to camera position and other factors, pedestrian video frames contain multiple angles of the pedestrian, and the gait features of the same person differ from different angles. Therefore, in the block convolution process, this invention not only convolves each input image but also concatenates two images in each group, and then performs block convolution on the concatenated features. This reduces the distance between features of the same person while increasing the distance between features of different people, making the features more discriminative and thus improving the model's recognition accuracy.

[0051] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0052] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0053] Figure 1 This is a flowchart illustrating a training method for a gait recognition network according to an exemplary embodiment.

[0054] Figure 2 This is a block convolutional schematic diagram of a gait recognition network according to an exemplary embodiment.

[0055] Figure 3 This is a schematic diagram of the main network portion of a gait recognition network according to an exemplary embodiment.

[0056] Figure 4 This is a block diagram illustrating a training system for a gait recognition network according to an exemplary embodiment.

[0057] Figure 5 This is a flowchart illustrating a gait recognition method according to an exemplary embodiment. Detailed Implementation

[0058] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of methods and systems consistent with some aspects of this application as detailed in the appended claims.

[0059] The above method uses a pedestrian binary contour map as input. However, in real-world scenarios, pedestrian postures vary greatly. When a pedestrian carries a backpack, wears a long coat, or carries other objects, it directly interferes with the pedestrian's contour features, leading to inaccurate gait recognition and affecting pedestrian verification. Furthermore, a single binary map does not contain the temporal features of pedestrian movement and is insufficient to distinguish gait changes between different people.

[0060] While RGB image sequences can extract appearance features, an additional LSTM (Long Short-Term Memory) network is required to extract gait variations over time, increasing computational overhead. Gait Energy Image (GEI), on the other hand, contains both pedestrian contour features and is derived from a sequence of binary images of the pedestrian over a period of time, making it a more suitable image format for pedestrian gait recognition. A key challenge in using GEI images is how to employ a lightweight network structure that uses the pedestrian's GEI image as input, remains unaffected by the appearance changes caused by backpacks or external objects, and accurately distinguishes between different individuals.

[0061] Figure 1 This is a flowchart illustrating a training method for a gait recognition network according to an exemplary embodiment. The method may include the following steps:

[0062] Step S1: Obtain M sets of training data. Each set of training data includes two gait energy maps, where M≥3.

[0063] Step S2: Divide the vertical height of the gait energy map into N parts, where N≥4;

[0064] Step S3: Simultaneously pass the N parts of the gait energy map through the basic convolutional layer to extract block features;

[0065] Step S4: The block features are fed into the second convolutional layer for processing. The second convolutional layer includes two branches. The first branch performs convolution on the block features of each image, and the second branch performs convolution on the concatenated block features of two images in a set.

[0066] Step S5: Construct a loss function based on the output of each convolutional layer, and use the loss function to supervise the training of the gait recognition network.

[0067] This application's scheme uses GEI images as network input, performs block convolution on each image after horizontal segmentation, and outputs the concatenated features. When a pedestrian moves, their leg state changes significantly, but when the legs are occluded, the features cannot represent the pedestrian's gait. Therefore, this scheme divides the image horizontally into multiple blocks. Even if the legs are occluded, the features of other parts remain distinguishable, thus improving the accuracy of pedestrian gait recognition. Furthermore, due to camera position and other factors, pedestrian video frames contain multiple angles of the pedestrian, and the gait features of the same person differ from different angles. Therefore, in the block convolution process, this invention not only convolves each input image but also concatenates two images in each group, and then performs block convolution on the concatenated features. This reduces the distance between features of the same person while increasing the distance between features of different people, making the features more discriminative and thus improving the model's recognition accuracy.

[0068] To further elaborate on the technical solution of this application, the features of the present invention will be briefly introduced first.

[0069] First, this invention uses pedestrian GEI images as input, with three sets of pedestrian GEI images input each time. To eliminate interference from external objects such as backpacks and long coats, this invention horizontally segments each image and employs block convolution (see reference). Figure 2 First, the invention extracts features of subtle changes in pedestrian details. Second, the input images are divided into three groups, with two images in each group. One group consists of images from the same person, and the other two groups consist of images from different people. This ensures that there are six different images in each input, and a five-fold loss function is constructed. Finally, the invention not only extracts the block features of each image but also concatenates the features of the two images in each group. The concatenated features are then subjected to block convolution again. The final output features are the result of weighted summation of the feature blocks, making the features more discriminative.

[0070] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0071] First, video streams are collected from the monitoring system. Pedestrian detection and segmentation are performed based on the video frames to obtain binary contour sequence images containing only pedestrians. Different pedestrians are numbered, with each pedestrian corresponding to a different ID number. Data from approximately 200 people are collected to create a pedestrian gait recognition dataset.

[0072] Based on the pedestrian binary contour map, the corresponding GEI map dataset is generated from the binary map of a sequence according to formula (1). The GEI map contains temporal information.

[0073]

[0074] Among them, I GEI Let GEI be the number of gait sequence images contained at that time point, and S be the number of images contained in the sequence. t Let x and y be the gait image at time t in the gait sequence, where x and y represent the coordinates in the image.

[0075] Six GEI images were randomly selected and fed into a convolutional neural network to extract features. Each pair of images was grouped together, with the first group consisting of images from the same ID number, and the second and third groups consisting of images from different ID numbers. It was ensured that one ID in the second group could not be the same as that in the first group. The aim was to construct a five-fold loss function. The input training data was constructed according to formula (3). Each image was horizontally divided into four parts and simultaneously passed through a basic convolutional layer to extract block features. Figure 2 As shown, T represents the number of input images, C, H, and W represent the number of channels, height, and width of the input training data, respectively, and Concat represents concatenation.

[0076] Four different ID numbers are randomly selected, denoted as id1, id2, id3, and id4; three gait energy maps are selected from id1, denoted as... For the remaining three ID numbers, select one gait energy map each, and record them as follows: Use the two gait energy maps of id1 as the first set of training data; group the remaining gait energy map of id1 with the gait energy maps of id2, id3, and id4 in pairs as the second and third sets of training data.

[0077] As mentioned above, the input data is defined as follows:

[0078] X = (x id1 ,x' id1 ,x” id1 ,x id2 ,x id3 ,x id4 (3)

[0079] Where, x id1 、x' id1 and This indicates that they all come from pedestrian image data labeled id1, and x id1 x id1 , There are three different images; x id2 ,x id3 ,x id4 These represent images from pedestrians id2, id3, and id4, respectively, making a total of 6 images in the input data. Here, we can stipulate that (x... id1 ,x' id1 (x) is the first pair of images in the input data.id1 ,x id2 (x) is the second group, (x) id3 ,x id4 ) is the third group.

[0080] The obtained features are then subjected to block convolution again. This time, the convolution has two branches: convolution on each individual image and concatenation of the convolutional features of two images within a set; for example... Figure 3 As shown, T represents the number of input images, and C, H, and W represent the number of channels, height, and width of the input training data, respectively.

[0081] After the convolutional layers described above, the summation is performed on each feature block. Since the arms and legs change the most when a pedestrian is walking, when there are situations such as carrying a backpack, wearing a long coat, or wearing a hat, the head and most of the legs are obscured, and these features no longer have discriminative power. Therefore, the second and fourth blocks, namely the pedestrian's shoulders, arms, and legs, are assigned larger weight values, and a weighted summation method is used to fuse each feature and output the fused feature.

[0082] After the convolution operation is completed, the features obtained from each branch are fed into the fully connected layer and mapped into a vector with the same dimension as the label. A five-fold recombination loss function L is constructed for supervised training, and L is defined by formula (8).

[0083] The five-fold loss function is summed from two terms. The first term is the triplet loss function L between samples. trip The purpose of the combination loss is to ensure that the distance between identical individuals is smaller than the distance between different individuals. The second term is the distance between the concatenated features of the first group and the concatenated features of the second group, minus the difference between the concatenated features of the second group and the concatenated features of the third group, representing the inter-class distance. The function of this five-fold combination loss function is to reduce intra-class distances and increase inter-class distances, thereby further distinguishing pedestrian gait.

[0084] Define the measurement calculation formula (4)

[0085]

[0086] d(x,y) represents the calculation of the distance between inputs x and y.

[0087] Triple loss function L trip Defined by formula (5):

[0088] L trip (A,P,N)=max(d(A,P)-d(A,N)+α,0) (5)

[0089] Where α represents the minimum feature interval distance.

[0090] Image feature loss function L imf Defined by formula (6):

[0091] L imf =η1L trip (x id1 ,x' id1 ,x id2 )+η2L trip (x id1 ,x' id1 ,x id3 )+η3L trip (x id1 ,x' id1 ,x id4 (6)

[0092] η1, η2, and η3 are configurable hyperparameters that represent the weight of the expression following them in the overall loss.

[0093] Image feature stitching feature loss function L conc The calculation is defined by formula (7):

[0094]

[0095] This represents the image stitching features of the first group. It's the second group. It's the third group.

[0096] The final loss function L is defined by equation (8):

[0097] L = η imf L imf +η conc L conc (8)

[0098] η imf It is the image feature loss function L imf The weighting parameter, η conc It is the concatenation feature loss function L conc The weight parameters.

[0099] Experiments have shown that the gait recognition method of this invention performs block convolution on images and weighted summation of block features to capture the detailed changes in pedestrian features. It can improve the accuracy of pedestrian gait recognition even when pedestrians are occluded by external objects such as carrying backpacks or wearing long coats, as well as when pedestrian posture changes from different camera angles. Furthermore, the block weighted features are used to construct a five-fold loss function to supervise the training of the convolutional neural network model, which further increases the distance between classes (between features of different people) and reduces the distance within classes (between features of the same person), thereby improving the accuracy of pedestrian gait recognition.

[0100] Figure 4 This is a block diagram illustrating a training system for a gait recognition network according to an exemplary embodiment. (Refer to...) Figure 4 The device includes:

[0101] The acquisition module 401 is used to acquire M sets of training data. Each set of training data includes two gait energy maps, where M≥3.

[0102] The segmentation module 402 is used to divide the vertical height of the gait energy map into N parts, where N≥4;

[0103] The basic convolutional module 403 is used to extract block features by simultaneously passing N parts of the gait energy map through the basic convolutional layer;

[0104] The second convolution module 404 is used to feed the block features into the second convolution layer for processing. The second convolution layer includes two branches: the first branch performs convolution on the block features of each image, and the second branch performs convolution on the concatenated block features of two images in a set.

[0105] The supervised training module 405 is used to construct a loss function based on the output of each convolutional layer, and to supervise the training of the gait recognition network through the loss function.

[0106] Regarding the system in the above embodiments, the specific steps for each module to perform operations have been described in detail in the embodiments related to the method, and will not be elaborated further here. Each module in the above-described gait recognition network training system can be implemented entirely or partially through software, hardware, or a combination thereof. Each module can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0107] Reference Figure 5 The embodiments of this application also provide a gait recognition method, including the following steps:

[0108] Retrieve camera video, perform pedestrian detection on video frames, and establish a pedestrian ID database;

[0109] For each video sequence with a given ID, target segmentation is performed to separate pedestrians and background.

[0110] During target segmentation, a sequence of binary contour maps of pedestrians is output, and the GEI map of the pedestrians is calculated based on the binary maps.

[0111] The GEI image of a pedestrian is input into a gait recognition network to extract gait features and establish a base database of pedestrian gait features; wherein, the gait recognition network is obtained by the training method of the gait recognition network described in any of the preceding embodiments;

[0112] When a pedestrian video frame is input for query, gait features are first extracted through a gait recognition network, and then similarity is measured with the features in the pedestrian gait feature database to find all video frames of the same person.

[0113] It is understood that the same or similar parts in the above embodiments can be referred to each other, and the contents not described in detail in some embodiments can be referred to the same or similar contents in other embodiments.

[0114] It should be noted that in the description of this application, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Furthermore, in the description of this application, unless otherwise stated, "a plurality of" means at least two.

[0115] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the function involved, as will be understood by those skilled in the art to which embodiments of this application pertain.

[0116] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0117] Those skilled in the art will understand that all or part of the steps of the methods described in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it includes one or a combination of the steps of the method embodiments.

[0118] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0119] The storage media mentioned above can be read-only memory, disk, or optical disk, etc.

[0120] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0121] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A training method for a gait recognition network, characterized in that, include: Obtain M sets of training data, each set consisting of two gait energy maps, where M ≥ 3; the steps for obtaining the training data include: Video streams are acquired from the monitoring system, and pedestrian detection and segmentation are performed based on video frames to obtain a binary contour sequence map containing only pedestrians. Different pedestrians are numbered, with each pedestrian corresponding to a different ID number, and each ID number corresponding to multiple binary contour sequence images; Generate the corresponding gait energy map based on the binary contour sequence map; When M is three, M sets of training data are obtained, including the following steps: Four different ID numbers are randomly selected and denoted as follows: id 1. id 2. id 3 and id 4; from Three gait energy maps were selected from the first three IDs, and one gait energy map was selected from each of the other three IDs. Will id Two gait energy maps from 1 were used as the first set of training data; Will The remaining gait energy map and , , The gait energy maps were grouped into pairs and used as training data for the second and third groups. Divide the gait energy map horizontally into N parts, where N≥4; The N parts of the gait energy map are simultaneously passed through a basic convolutional layer to extract block features; The block features are fed into the second convolutional layer for processing. The second convolutional layer includes two branches: the first branch performs convolution on the block features of each image, and the second branch performs convolution on the concatenated block features of two images in a set. A loss function is constructed based on the output of each convolutional layer, and the gait recognition network is trained under supervision using the loss function. The loss function is constructed based on the output of each convolutional layer, including: constructing a five-recombination loss function, which is the sum of two terms; where the first term is: the triplet loss function among the three sets of training data. L trip The combined loss is: the distance between the concatenated features of the first set of training data and the concatenated features of the second set of training data, minus the distance between the concatenated features of the second set of training data and the concatenated features of the third set of training data.

2. The method according to claim 1, characterized in that, Triple loss function L trip for: ； in, α Indicates the minimum interval distance of the features; This indicates that the distance between x and y is calculated.

3. The method according to claim 2, characterized in that, Image feature loss function L imf for: ； in , , It is a configurable hyperparameter that represents the weight of the expression multiplied with it in the overall loss.

4. The method according to claim 3, characterized in that, Image feature stitching feature loss function L conc for: ； This represents the image stitching features of the first set of training data. These are the image stitching features from the second set of training data. These are the image stitching features of the third set of training data.

5. The method according to claim 4, characterized in that, Five-fold recombination loss function L for: ； in It is an image feature loss function L imf The weight parameters, It is a concatenation feature loss function L conc The weight parameters.

6. A training system for a gait recognition network, characterized in that, include: The acquisition module is used to acquire M sets of training data. Each set of training data includes two gait energy maps, where M ≥ 3. The steps for acquiring the training data include: Video streams are acquired from the monitoring system, and pedestrian detection and segmentation are performed based on video frames to obtain a binary contour sequence map containing only pedestrians. Different pedestrians are numbered, with each pedestrian corresponding to a different ID number, and each ID number corresponding to multiple binary contour sequence images; Generate the corresponding gait energy map based on the binary contour sequence map; When M is three, M sets of training data are obtained, including the following steps: Four different ID numbers are randomly selected and denoted as follows: id 1. id 2. id 3 and id 4; from Three gait energy maps were selected from the first three IDs, and one gait energy map was selected from each of the other three IDs. Will id Two gait energy maps from 1 were used as the first set of training data; Will The remaining gait energy map and , , The gait energy maps were grouped into pairs and used as training data for the second and third groups. The segmentation module is used to horizontally segment the gait energy map into N parts, where N≥4; The basic convolutional module is used to extract block features by simultaneously passing N parts of the gait energy map through the basic convolutional layer; The second convolution module is used to feed the block features into the second convolution layer for processing. The second convolution layer includes two branches: the first branch performs convolution on the block features of each image, and the second branch performs convolution on the concatenated block features of two images in a set. The supervised training module is used to construct a loss function based on the output of each convolutional layer, and to supervise the training of the gait recognition network using this loss function. Specifically, it is used to construct a five-fold recombination loss function, which is the sum of two terms. The first term is the triplet loss function between the three sets of training data. L trip The combined loss is: the distance between the concatenated features of the first set of training data and the concatenated features of the second set of training data, minus the distance between the concatenated features of the second set of training data and the concatenated features of the third set of training data.

7. A gait recognition method, characterized in that, include: Retrieve camera video, perform pedestrian detection on video frames, and establish a pedestrian ID database; For each video sequence with a given ID, target segmentation is performed to separate pedestrians and background. During target segmentation, a sequence of binary contour maps of pedestrians is output, and the GEI map of the pedestrians is calculated based on the binary maps. The GEI image of a pedestrian is input into a gait recognition network to extract gait features and establish a basic database of pedestrian gait features. When a pedestrian video frame to be queried is input, gait features are first extracted through a gait recognition network, and then similarity is measured with the features in the pedestrian gait feature database to find all video frames of the same person. The gait recognition network is obtained by the training method of the gait recognition network according to any one of claims 1-5.

Images