Gait recognition method, device and equipment and storage medium

Abstract

The invention discloses a gait recognition method and device, equipment and a storage medium, and relates to the technical field of biological recognition. Comprising the following steps: extracting a pedestrian image from a to-be-recognized video sequence by using a preset image processing technology to obtain a target gait sequence, and performing time sequence alignment processing on the target gait sequence based on a dynamic time normalization technology to generate a gait energy diagram; inputting the gait energy map into a preset deep neural network convolution model to extract space attention features and channel attention features, and performing feature fusion to obtain an enhanced feature map; performing feature integration on the enhanced feature map to obtain an integrated vector, and mapping the integrated vector to a preset identity output tag space to obtain an identity prediction output vector; and inputting the identity prediction output vector into a preset classifier for training to obtain corresponding gait probability distribution, and determining a corresponding gait recognition prediction result according to the gait probability distribution. In this way, the accuracy and efficiency of gait recognition can be improved.

Description

Technical Field

[0001] The present invention relates to the field of biometric identification technology, and in particular to a gait recognition method, device, equipment and storage medium. Background Art

[0002] Gait recognition technology, as an identity authentication method based on individual walking posture, has received widespread attention in recent years. Unlike fingerprint or iris recognition, gait recognition does not require the active cooperation of individuals and can be performed without the knowledge of individuals or at a distance. This feature makes it show unique application value in the fields of monitoring and security verification. Gait recognition technology mainly relies on walking sequences captured by video surveillance systems. By analyzing the dynamic characteristics of individuals in these sequences, such as step length, speed, body swing, etc., the gait recognition system can identify or verify the identity of individuals. Early gait recognition research focused on manually extracting features from video sequences and then using these features for individual identification. As a result, when processing large-scale data, high computing resources are often required, which limits their application in real-time monitoring systems. Therefore, how to improve the efficiency and accuracy of gait recognition technology is a problem that needs to be solved. Summary of the invention

[0003] In view of this, the purpose of the present invention is to provide a gait recognition method, device, equipment and storage medium, which can improve the efficiency and accuracy of gait recognition technology. The specific scheme is as follows:

[0004] In a first aspect, the present application discloses a gait recognition method, comprising:

[0005] Extracting pedestrian images from the video sequence to be identified by using a preset image processing technology to obtain a target gait sequence, and performing time series alignment processing on the target gait sequence based on a dynamic time normalization technology to generate a gait energy graph;

[0006] Inputting the gait energy map into a preset deep neural network convolution model to extract spatial attention features and channel attention features in the gait energy map, and performing feature fusion on the spatial attention features and the channel attention features to obtain an enhanced feature map;

[0007] Performing feature integration on the enhanced feature map to obtain an integrated vector, and mapping the integrated vector to a preset identity output label space to obtain an identity prediction output vector;

[0008] The identity prediction output vector is input into a preset classifier for training to obtain a corresponding gait probability distribution, and a corresponding gait recognition prediction result is determined according to the gait probability distribution.

[0009] Optionally, the extracting pedestrian images from the to-be-identified video sequence using a preset image processing technology to obtain a target gait sequence includes:

[0010] A video sequence to be identified is obtained, and pedestrian images are segmented from the video sequence to be identified using background difference technology, shadow removal technology and dynamic threshold processing technology to obtain a target gait sequence.

[0011] Optionally, performing time series alignment processing on the target gait sequence based on the dynamic time normalization technology to generate a gait energy graph includes:

[0012] The minimum distance between each gait time series in the target gait sequence is calculated by a preset recursive formula to obtain the best matching path between the gait time series, and the gait time series are aligned based on the minimum cumulative distance to obtain the gait energy graph corresponding to the target gait sequence.

[0013] Optionally, the step of inputting the gait energy map into a preset deep neural network convolution model to extract spatial attention features and channel attention features in the gait energy map includes:

[0014] A preset deep neural network convolution model is used to extract the target features in the gait energy map, and the spatial attention mechanism and channel attention mechanism in the preset deep neural network convolution model are used to process the target features respectively to obtain the spatial attention features and channel attention features in the gait energy map.

[0015] Optionally, the using of the spatial attention mechanism and the channel attention mechanism in the preset deep neural network convolution model to process the target features respectively to obtain the spatial attention features and the channel attention features in the gait energy graph includes:

[0016] Determine the spatial weight of the target feature by using the sigmoid activation function through the spatial attention mechanism in the preset deep neural network convolution model to obtain the spatial attention feature;

[0017] The channel attention mechanism in the preset deep neural network convolution model uses global average pooling and a fully connected layer to determine the channel weight of the target feature to obtain the channel attention feature.

[0018] Optionally, before inputting the identity prediction output vector into a preset classifier for training to obtain a corresponding gait probability distribution, the method further includes:

[0019] Inputting a preset training classification function into an initial classifier to obtain a predicted gait probability distribution corresponding to each preset identity tag, and determining a probability difference between the predicted gait probability distribution corresponding to each preset identity tag and a preset true probability based on a preset cross entropy loss function;

[0020] The gradient of the initial classifier is determined by using a preset back-propagation algorithm and the probability difference, and the model parameters of the initial classifier are updated by using a preset optimization algorithm and the gradient to obtain the preset classifier.

[0021] Optionally, inputting the identity prediction output vector into a preset classifier for training to obtain a corresponding gait probability distribution, and determining a corresponding gait recognition prediction result according to the gait probability distribution, includes:

[0022] Input the identity prediction output vector into a preset classifier for training to obtain a corresponding activation value, and use a Sigmoid activation function to convert the activation value into a gait probability distribution;

[0023] The identity tag with the highest probability value in the gait probability distribution is used to determine the corresponding gait recognition prediction result.

[0024] In a second aspect, the present application discloses a gait recognition device, comprising:

[0025] A gait energy map generation module is used to extract pedestrian images from the video sequence to be identified using a preset image processing technology to obtain a target gait sequence, and perform time series alignment processing on the target gait sequence based on a dynamic time normalization technology to generate a gait energy map;

[0026] A feature extraction module, used for inputting the gait energy map into a preset deep neural network convolution model to extract spatial attention features and channel attention features in the gait energy map, and performing feature fusion on the spatial attention features and the channel attention features to obtain an enhanced feature map;

[0027] A vector mapping module, used for performing feature integration on the enhanced feature map to obtain an integrated vector, and mapping the integrated vector to a preset identity output label space to obtain an identity prediction output vector;

[0028] The gait recognition module is used to input the identity prediction output vector into a preset classifier for training to obtain a corresponding gait probability distribution, and determine a corresponding gait recognition prediction result according to the gait probability distribution.

[0029] In a third aspect, the present application discloses an electronic device, comprising:

[0030] Memory, used to store computer programs;

[0031] A processor is used to execute the computer program to implement the aforementioned gait recognition method.

[0032] In a fourth aspect, the present application discloses a computer-readable storage medium for storing a computer program, wherein the computer program implements the aforementioned gait recognition method when executed by a processor.

[0033] It can be seen that in the present application, a preset image processing technology is used to extract pedestrian images from a video sequence to be identified to obtain a target gait sequence, and the target gait sequence is subjected to time series alignment processing based on a dynamic time normalization technology to generate a gait energy map; the gait energy map is input into a preset deep neural network convolution model to extract spatial attention features and channel attention features in the gait energy map, and the spatial attention features and the channel attention features are feature fused to obtain an enhanced feature map; feature integration is performed on the enhanced feature map to obtain an integrated vector, and the integrated vector is mapped to a preset identity output label space to obtain an identity prediction output vector; the identity prediction output vector is input into a preset classifier for training to obtain a corresponding gait probability distribution, and the corresponding gait recognition prediction result is determined according to the gait probability distribution. In this way, by using dynamic time normalization technology to time-align gait sequences, gait sequences of different lengths can be effectively processed and the generated gait energy map can be ensured to accurately reflect the dynamic characteristics of individual gait; by extracting the spatial attention features and channel attention features in the gait energy map, the most important gait feature areas for recognition can be automatically selected to improve recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0035] Figure 1 A flow chart of a gait recognition method disclosed in this application;

[0036] Figure 2 A schematic diagram of a specific neural network model structure disclosed in this application;

[0037] Figure 3 A specific attention feature extraction flow chart disclosed in this application;

[0038] Figure 4 A schematic diagram of a specific gait recognition result disclosed in this application;

[0039] Figure 5 This is a schematic diagram of the structure of a gait recognition device disclosed in this application;

[0040] Figure 6 This is a structural diagram of an electronic device disclosed in this application. DETAILED DESCRIPTION

[0041] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0042] With the development of machine learning and deep learning technologies, a series of new methods based on these technologies have emerged in the field of gait recognition. These methods usually include the following steps: first, extract the individual's profile or profile sequence from the video sequence; then, use a deep learning model, such as a convolutional neural network (CNN), to automatically extract gait features from the profile sequence; finally, use the extracted features to identify or verify the individual. However, this type of gait recognition method often requires higher computing resources when processing large-scale data. Therefore, this application will specifically introduce a gait recognition method that can improve the accuracy and efficiency of gait recognition.

[0043] See also Figure 1 As shown, the embodiment of the present application discloses a gait recognition method, comprising:

[0044] Step S11: extract pedestrian images from the video sequence to be identified using a preset image processing technology to obtain a target gait sequence, and perform time series alignment processing on the target gait sequence based on a dynamic time normalization technology to generate a gait energy graph.

[0045] In this embodiment, data preprocessing is the first key step to achieve high-precision gait recognition. The method of extracting pedestrian images from the video sequence to be identified by using a preset image processing technology to obtain a target gait sequence includes: obtaining a video sequence to be identified, and segmenting pedestrian images from the video sequence to be identified by using background difference technology, shadow removal technology and dynamic threshold processing technology to obtain a target gait sequence. In other words, firstly, the pedestrian image is automatically segmented from the video sequence by using background difference, shadow removal and dynamic threshold processing. Then, the minimum distance between each gait time sequence in the target gait sequence is calculated by a preset recursive formula to obtain the best matching path between each gait time sequence, and the alignment processing of each gait time sequence is performed based on the minimum cumulative distance to obtain the gait energy map corresponding to the target gait sequence. That is, the pedestrian gait sequence is processed by the dynamic time normalization (DTW, i.e. Dynamic Time Warping) technology, so that gait sequences of different time lengths are aligned to the same time frame, thereby generating a gait energy map (GEI, i.e. Gait Engery Image). That is to say, the gait sequence is aligned through the dynamic time normalization (DTW) technology to generate a standardized gait energy graph. It should be noted here that dynamic time normalization is a technology used to align time series of different lengths, especially suitable for inconsistent gait cycles in gait recognition. DTW achieves alignment by calculating the minimum cumulative distance between two time series. Its recursive formula is:

[0046] ;

[0047] Among them, DTW(i, j) represents the cumulative value of the matching cost or distance between the i-th element and the j-th element in two time series, and d(i, j) represents the distance between the i-th and j-th elements in the time series, which is calculated using the Euclidean distance. DTW uses dynamic programming to find the best matching path between two sequences, thereby aligning sequences with different phase cycles. In this way, by introducing dynamic time normalization and attention mechanism, the adaptability to different perspectives is improved; and through efficient image processing and feature extraction technology, the recognition accuracy under various environmental conditions is improved; through improved data preprocessing steps, the model's processing ability for low-resolution videos is enhanced, and the recognition accuracy is improved.

[0048] Step S12: input the gait energy map into a preset deep neural network convolution model to extract the spatial attention features and channel attention features in the gait energy map, and perform feature fusion on the spatial attention features and the channel attention features to obtain an enhanced feature map.

[0049] In this embodiment, Figure 2As shown, the present application adopts an improved deep convolutional neural network model, which is specially optimized for the characteristics of GEI. The model structure includes a convolution layer, a batch normalization layer, an activation function, a maximum pooling layer and a random drop layer. By introducing the attention mechanism, the model can focus on the key areas in GEI and improve the accuracy of recognition. The optimized CNN model and efficient classification algorithm reduce the demand for computing resources and are suitable for real-time applications. Among them, the feature extraction module adopts an improved deep convolutional neural network structure. The network contains three main convolution blocks, each of which consists of a convolution layer, a batch normalization layer, a ReLU activation function (Linear rectification function) and a maximum pooling layer. In particular, an attention module is introduced after each convolution block to enhance the ability to extract key features. As the number of network layers increases, the spatial dimension of the feature map gradually decreases, while the number of channels gradually increases, realizing the effective extraction from low-level features to high-level features. The convolution layer is the core component of CNN and is used to extract features from the input image. The convolution operation performs a dot product operation on the convolution kernel and the local area of ​​the input image by means of a sliding window. The formula is:

[0050] ;

[0051] Among them, M and N are the height and width of the convolution kernel, respectively, m and n are the indices of the convolution kernel, indicating the position of the convolution kernel in the input image or the element of the local area, y i,j Output the value of the (i, j)th position of the image, W m,n is the weight value of the convolution kernel (filter) at position (m, n), X i+m，j+n The pixel value x of the input image at position (i+m, j+n) is the input image, w is the convolution kernel, and b is the bias. The convolution operation can effectively extract low-level features such as edges and textures in the image. Batch normalization is used to accelerate the training process of the neural network and improve the stability of the model. It normalizes each small batch of data, and the formula is:

[0052] ;

[0053] in, is the output after batch normalization, is the input value of the kth sample in the mini-batch, and are the mean and variance of the mini-batch, is a small constant used to prevent division by zero errors. Batch normalization helps alleviate the gradient vanishing and gradient exploding problems. ReLU is a commonly used activation function in deep learning to introduce nonlinear factors. It is defined as:

[0054] ;

[0055] The ReLU function can effectively alleviate the gradient vanishing problem and accelerate the convergence speed of the network. MaxPooling is used to reduce the dimension of the feature map while retaining important feature information. It applies a pooling window on the feature map and selects the maximum value in the window. The formula is:

[0056] ;

[0057] Among them, the maximum pooling helps to reduce the amount of calculation and improve the spatial invariance of the model. Dropout is a regularization technique used to prevent overfitting of neural networks. It is achieved by randomly dropping neurons. The formula is:

[0058] ;

[0059] Where p is the dropout rate. Dropout improves the generalization ability of the model by reducing the interdependence between neurons. The attention mechanism emphasizes important features and suppresses unimportant parts through the weight coefficients obtained through learning. Figure 3 As shown, the step of inputting the gait energy map into a preset deep neural network convolution model to extract the spatial attention features and channel attention features in the gait energy map includes: extracting the target features in the gait energy map using the preset deep neural network convolution model, and processing the target features using the spatial attention mechanism and channel attention mechanism in the preset deep neural network convolution model to obtain the spatial attention features and channel attention features in the gait energy map. Specifically, the spatial weight of the target feature is determined by the spatial attention mechanism in the preset deep neural network convolution model using the Sigmoid activation function to obtain the spatial attention feature; the channel weight of the target feature is determined by the channel attention mechanism in the preset deep neural network convolution model using global average pooling and a fully connected layer to obtain the channel attention feature. It should be noted that the attention mechanism module includes two branches: spatial attention and channel attention. The spatial attention branch calculates the spatial weight by 1×1 convolution and Sigmoid activation function, while the channel attention branch calculates the channel weight by global average pooling and a fully connected layer. The outputs of the two branches are feature fused to generate an enhanced feature map, which effectively improves the model's attention to key areas. Specifically, it is achieved by calculating the importance weight of the feature, the formula is:

[0060] ;

[0061] Among them, T is the total number of features, The attention weight of feature i. and is the index of the feature in this formula, indicating the calculation of the weighted importance of each feature, where i is the index of the feature or element currently being calculated, indicating a feature of interest, j is the index used for the normalization process, indicating the accumulation of the score of each feature among all features, and e i The importance of features is calculated by a learnable function. The attention mechanism can increase the model's attention to key features, thereby improving the accuracy of recognition. In this way, it can focus on key gait features that are not obscured, improving the recognition performance under occlusion.

[0062] Step S13: performing feature integration on the enhanced feature map to obtain an integrated vector, and mapping the integrated vector to a preset identity output label space to obtain an identity prediction output vector.

[0063] In this embodiment, after being processed by layers such as convolution, batch normalization, activation, and pooling, the feature map is flattened into a one-dimensional vector. This step is to convert the multidimensional feature map into a format suitable for processing by the fully connected layer. Assume that the size of the feature map is C (Channel) × H (Height) × W (Width), and the length of the flattened feature vector is C × H × W. Then, the flattened feature vector is mapped to the output label space through the fully connected layer. Assuming that there are N categories (i.e., N different identity labels), the output of the fully connected layer is an N-dimensional vector. The calculation formula of the fully connected layer is:

[0064] ;

[0065] Among them, x i is the i-th element of the flattened eigenvector, w ji is the element of the weight matrix, b j is the jth element of the bias vector, z j is the jth element of the output vector, and M is the length of the flattened feature vector.

[0066] Step S14: inputting the identity prediction output vector into a preset classifier for training to obtain a corresponding gait probability distribution, and determining a corresponding gait recognition prediction result according to the gait probability distribution.

[0067] In this embodiment, before the identity prediction output vector is input into a preset classifier for training to obtain the corresponding gait probability distribution, it also includes: inputting a preset training classification function into the initial classifier to obtain the predicted gait probability distribution corresponding to each preset identity tag, and determining the probability difference between the predicted gait probability distribution corresponding to each preset identity tag and the preset true probability based on a preset cross entropy loss function; determining the gradient of the initial classifier using a preset back propagation algorithm and the probability difference, and updating the model parameters of the initial classifier using a preset optimization algorithm and the gradient to obtain the preset classifier. First of all, it should be noted that the Softmax classifier is used to convert the output vector (z) of the fully connected layer into a probability distribution. The formula of the Softmax function is:

[0068] ;

[0069] Among them, z is the output vector of the fully connected layer, which contains the unnormalized category scores (logits) of the network, is the input score z j exponential function. j is the index of Softmax output, indicating the index of the category currently calculated. It represents the result of converting the score (logit) of category k through the exponential function. k is the index in the exponential normalization process, which represents all categories used for normalization. Represents the probability that the input belongs to the jth category. The Softmax function ensures that the sum of the output probability distribution is 1, so that each output value can be interpreted as the confidence that it belongs to a certain category. During model training, the cross entropy loss function is used to measure the difference between the probability distribution predicted by the model and the actual label. The formula for the cross entropy loss function is:

[0070] ;

[0071] Among them, y j is the indicator function of the actual label (one-hot encoding), is the predicted probability. By minimizing the cross entropy loss, the model is able to more accurately predict the individual's gait category. During training, the backpropagation algorithm is used to calculate the gradient of the loss function with respect to the model parameters. This process involves the chain rule to calculate the gradient layer by layer. Assume that the loss function is (L) and the model parameters are ( ), then the gradient calculation formula is:

[0072] ;

[0073] Among them, optimization algorithms (such as stochastic gradient descent, SGD) use these gradients to update the model parameters to minimize the loss function, The meaning is that the loss function L has an effect on the model parameters The gradient here is a vector containing the partial derivatives of the loss function with respect to all model parameters. By calculating this gradient, the back propagation algorithm can guide how to adjust the parameters of the model to minimize the loss function. The parameter update formula of SGD is:

[0074] ;

[0075] in, is the learning rate, which determines the step size of each update.

[0076] In this embodiment, the identity prediction output vector is input into a preset classifier for training to obtain a corresponding gait probability distribution, and the corresponding gait recognition prediction result is determined according to the gait probability distribution, including: inputting the identity prediction output vector into a preset classifier for training to obtain a corresponding activation value, and using a Sigmoid activation function to convert the activation value into a gait probability distribution; determining the corresponding gait recognition prediction result with the identity label with the highest probability value in the gait probability distribution. That is, after the model training is completed, when a new gait energy graph (GEI) is input, the model calculates the probability distribution of each category through forward propagation. The process of forward propagation is to pass the input data through the network layer by layer, and finally obtain the activation value of the output layer, and then the output layer usually uses a Softmax function to convert the activation value into a probability distribution. The final classification result is the category with the highest probability, that is:

[0077] ;

[0078] That is, after obtaining the probabilities corresponding to each identity label, the category with the highest probability is selected as the prediction result of the model. After obtaining the final prediction result, the recognition result is displayed through a user-friendly interface, such as Figure 4 As shown, it includes identity information, recognition confidence score and historical gait data comparison results.

[0079] It can be seen that in this embodiment, a preset image processing technology is used to extract pedestrian images from a video sequence to be identified to obtain a target gait sequence, and the target gait sequence is subjected to time series alignment processing based on a dynamic time normalization technology to generate a gait energy map; the gait energy map is input into a preset deep neural network convolution model to extract spatial attention features and channel attention features in the gait energy map, and feature fusion is performed on the spatial attention features and the channel attention features to obtain an enhanced feature map; feature integration is performed on the enhanced feature map to obtain an integrated vector, and the integrated vector is mapped to a preset identity output label space to obtain an identity prediction output vector; the identity prediction output vector is input into a preset classifier for training to obtain a corresponding gait probability distribution, and the corresponding gait recognition prediction result is determined according to the gait probability distribution. In this way, by using dynamic time normalization technology to time-align gait sequences, gait sequences of different lengths can be effectively processed and the generated gait energy map can be ensured to accurately reflect the dynamic characteristics of individual gait; by extracting the spatial attention features and channel attention features in the gait energy map, the most important gait feature areas for recognition can be automatically selected to improve recognition accuracy.

[0080] refer to Figure 5 The present application also discloses a gait recognition device, including:

[0081] A gait energy map generation module 11 is used to extract pedestrian images from the video sequence to be identified using a preset image processing technology to obtain a target gait sequence, and perform time series alignment processing on the target gait sequence based on a dynamic time normalization technology to generate a gait energy map;

[0082] A feature extraction module 12 is used to input the gait energy map into a preset deep neural network convolution model to extract the spatial attention features and channel attention features in the gait energy map, and perform feature fusion on the spatial attention features and the channel attention features to obtain an enhanced feature map;

[0083] A vector mapping module 13, configured to perform feature integration on the enhanced feature map to obtain an integrated vector, and map the integrated vector to a preset identity output label space to obtain an identity prediction output vector;

[0084] The gait recognition module 14 is used to input the identity prediction output vector into a preset classifier for training to obtain a corresponding gait probability distribution, and determine a corresponding gait recognition prediction result according to the gait probability distribution.

[0085] It can be seen that in this embodiment, by using the dynamic time normalization technology to time-align the gait sequence, gait sequences of different lengths can be effectively processed, and it is ensured that the generated gait energy map accurately reflects the dynamic characteristics of the individual gait; by extracting the spatial attention features and channel attention features in the gait energy map, the gait feature areas that are most important for identification can be automatically selected to improve the recognition accuracy.

[0086] In some specific embodiments, the gait energy graph generating module 11 may specifically include:

[0087] The gait graph extraction unit is used to obtain a video sequence to be identified, and segment pedestrian images from the video sequence to be identified using background difference technology, shadow removal technology and dynamic threshold processing technology to obtain a target gait sequence.

[0088] In some specific embodiments, the gait energy graph generating module 11 may specifically include:

[0089] A gait graph processing unit is used to calculate the minimum distance between each gait time series in the target gait sequence through a preset recursive formula to obtain the best matching path between the gait time series, and align the gait time series based on the minimum cumulative distance to obtain a gait energy graph corresponding to the target gait sequence.

[0090] In some specific embodiments, the feature extraction module 12 may specifically include:

[0091] A feature extraction submodule is used to extract target features in the gait energy map using a preset deep neural network convolution model, and to process the target features using the spatial attention mechanism and channel attention mechanism in the preset deep neural network convolution model to obtain spatial attention features and channel attention features in the gait energy map.

[0092] In some specific embodiments, the feature extraction submodule may specifically include:

[0093] A spatial attention feature extraction unit, used to determine the spatial weight of the target feature by using a sigmoid activation function through a spatial attention mechanism in the preset deep neural network convolution model to obtain a spatial attention feature;

[0094] A channel attention feature extraction unit is used to determine the channel weight of the target feature by using the channel attention mechanism in the preset deep neural network convolution model using global average pooling and a fully connected layer to obtain a channel attention feature.

[0095] In some specific embodiments, the gait recognition device may further include:

[0096] A probability calculation module, used to input a preset training classification function into an initial classifier to obtain a predicted gait probability distribution corresponding to each preset identity tag, and determine a probability difference between the predicted gait probability distribution corresponding to each preset identity tag and a preset true probability based on a preset cross entropy loss function;

[0097] The classifier optimization module is used to determine the gradient of the initial classifier using a preset back-propagation algorithm and the probability difference, and to update the model parameters of the initial classifier using a preset optimization algorithm and the gradient to obtain the preset classifier.

[0098] In some specific embodiments, the gait recognition module 14 may specifically include:

[0099] An activation value determination unit, used for inputting the identity prediction output vector into a preset classifier for training to obtain a corresponding activation value, and converting the activation value into a gait probability distribution using a Sigmoid activation function;

[0100] The gait recognition unit is used to determine the corresponding gait recognition prediction result according to the identity tag with the highest probability value in the gait probability distribution.

[0101] Furthermore, the present application also discloses an electronic device. Figure 6 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content in the diagram cannot be regarded as any limitation on the scope of use of the present application.

[0102] Figure 6 The present invention provides a schematic diagram of the structure of an electronic device 20 provided in an embodiment of the present application. The electronic device 20 may specifically include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 is used to store a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the gait recognition method disclosed in any of the aforementioned embodiments. In addition, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0103] In this embodiment, the power supply 23 is used to provide working voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and the external device, and the communication protocol it follows is any communication protocol that can be applied to the technical solution of the present application, and is not specifically limited here; the input and output interface 25 is used to obtain external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs and is not specifically limited here.

[0104] In addition, the memory 22, as a carrier for storing resources, can be a read-only memory, a random access memory, a disk or an optical disk, etc. The resources stored thereon can include an operating system 221, a computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0105] The operating system 221 is used to manage and control the hardware devices and computer program 222 on the electronic device 20, and can be Windows Server, Netware, Unix, Linux, etc. In addition to the computer program that can be used to complete the gait recognition method performed by the electronic device 20 disclosed in any of the aforementioned embodiments, the computer program 222 can further include computer programs that can be used to complete other specific tasks.

[0106] Furthermore, the present application also discloses a computer-readable storage medium for storing a computer program; wherein the computer program, when executed by a processor, implements the gait recognition method disclosed above. For the specific steps of the method, reference may be made to the corresponding contents disclosed in the above embodiments, and no further description will be given here.

[0107] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0108] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0109] The steps of the method or algorithm described in conjunction with the embodiments disclosed herein may be implemented directly using hardware, a software module executed by a processor, or a combination of the two. The software module may be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0110] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0111] The technical solution provided by the present application is introduced in detail above. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea. At the same time, for general technicians in this field, according to the idea of ​​the present application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. A gait recognition method, characterized in that: include: Extracting pedestrian images from the video sequence to be identified by using a preset image processing technology to obtain a target gait sequence, and performing time series alignment processing on the target gait sequence based on a dynamic time normalization technology to generate a gait energy graph; Inputting the gait energy map into a preset deep neural network convolution model to extract spatial attention features and channel attention features in the gait energy map, and performing feature fusion on the spatial attention features and the channel attention features to obtain an enhanced feature map; Performing feature integration on the enhanced feature map to obtain an integrated vector, and mapping the integrated vector to a preset identity output label space to obtain an identity prediction output vector; The identity prediction output vector is input into a preset classifier for training to obtain a corresponding gait probability distribution, and a corresponding gait recognition prediction result is determined according to the gait probability distribution.

2. The gait recognition method according to claim 1, characterized in that: The method of extracting pedestrian images from a video sequence to be identified using a preset image processing technology to obtain a target gait sequence includes: A video sequence to be identified is obtained, and pedestrian images are segmented from the video sequence to be identified using background difference technology, shadow removal technology and dynamic threshold processing technology to obtain a target gait sequence.

3. The gait recognition method according to claim 1, characterized in that: The step of performing time series alignment processing on the target gait sequence based on the dynamic time normalization technology to generate a gait energy graph includes: The minimum distance between each gait time series in the target gait sequence is calculated by a preset recursive formula to obtain the best matching path between the gait time series, and the gait time series are aligned based on the minimum cumulative distance to obtain the gait energy graph corresponding to the target gait sequence.

4. The gait recognition method according to claim 1, characterized in that: The step of inputting the gait energy graph into a preset deep neural network convolution model to extract spatial attention features and channel attention features in the gait energy graph comprises: A preset deep neural network convolution model is used to extract the target features in the gait energy map, and the spatial attention mechanism and channel attention mechanism in the preset deep neural network convolution model are used to process the target features respectively to obtain the spatial attention features and channel attention features in the gait energy map.

5. The gait recognition method according to claim 4, characterized in that: The using of the spatial attention mechanism and the channel attention mechanism in the preset deep neural network convolution model to process the target features respectively to obtain the spatial attention features and the channel attention features in the gait energy graph includes: Determine the spatial weight of the target feature by using the sigmoid activation function through the spatial attention mechanism in the preset deep neural network convolution model to obtain the spatial attention feature; The channel attention mechanism in the preset deep neural network convolution model uses global average pooling and a fully connected layer to determine the channel weight of the target feature to obtain the channel attention feature.

6. The gait recognition method according to claim 1, characterized in that: Before inputting the identity prediction output vector into a preset classifier for training to obtain a corresponding gait probability distribution, the method further includes: Inputting a preset training classification function into an initial classifier to obtain a predicted gait probability distribution corresponding to each preset identity tag, and determining a probability difference between the predicted gait probability distribution corresponding to each preset identity tag and a preset true probability based on a preset cross entropy loss function; The gradient of the initial classifier is determined by using a preset back-propagation algorithm and the probability difference, and the model parameters of the initial classifier are updated by using a preset optimization algorithm and the gradient to obtain the preset classifier.

7. The gait recognition method according to any one of claims 1 to 6, characterized in that: The inputting the identity prediction output vector into a preset classifier for training to obtain a corresponding gait probability distribution, and determining a corresponding gait recognition prediction result according to the gait probability distribution, comprises: Input the identity prediction output vector into a preset classifier for training to obtain a corresponding activation value, and use a Sigmoid activation function to convert the activation value into a gait probability distribution; The identity tag with the highest probability value in the gait probability distribution is used to determine the corresponding gait recognition prediction result.

8. A gait recognition device, characterized in that: include: A gait energy map generation module is used to extract pedestrian images from the video sequence to be identified using a preset image processing technology to obtain a target gait sequence, and perform time series alignment processing on the target gait sequence based on a dynamic time normalization technology to generate a gait energy map; A feature extraction module, used for inputting the gait energy map into a preset deep neural network convolution model to extract spatial attention features and channel attention features in the gait energy map, and performing feature fusion on the spatial attention features and the channel attention features to obtain an enhanced feature map; A vector mapping module, used for performing feature integration on the enhanced feature map to obtain an integrated vector, and mapping the integrated vector to a preset identity output label space to obtain an identity prediction output vector; The gait recognition module is used to input the identity prediction output vector into a preset classifier for training to obtain a corresponding gait probability distribution, and determine a corresponding gait recognition prediction result according to the gait probability distribution.

9. An electronic device, characterized in that: include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the gait recognition method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: Used to store a computer program, which, when executed by a processor, implements the gait recognition method according to any one of claims 1 to 7.

Images