A pedestrian identity recognition method, system and storage medium

Abstract

The present application relates to the technical field of image recognition, and discloses a pedestrian identity recognition method and system and a storage medium, which combine gait contour image features and dense optical flow image features, and adopt an adaptive weighting method to extract a SIL-FLOW feature map of a pedestrian, thereby solving the problem that only using gait contour images can only capture contour edge information and ignore motion information inside the contour, improving recognition accuracy; a gait feature coding network is used to replace a traditional convolutional neural network, and input fused feature maps are converted into low-resolution, high-semantic feature maps, so that the output image contains gait semantic features highly related to the edge of a pedestrian, further improving recognition accuracy; the gait feature encoder introduces a residual branch 3, which better alleviates the gradient vanishing and explosion problems of traditional deep neural networks, and eliminates the interference of irrelevant information in the feature map, further improving the accuracy of pedestrian identity recognition.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to a pedestrian identification method, system, and storage medium. Background Technology

[0002] Biometric-based pedestrian identification methods mainly consist of technologies such as facial recognition, fingerprint recognition, iris recognition, and gait feature recognition. Gait refers to a pedestrian's walking posture, including the swinging motion of the legs and arms, and is an important characteristic of pedestrian behavior. Compared with other biometric identification technologies, gait feature recognition technology has significant advantages such as being non-contact, requiring lower resolution of the images to be identified, and being difficult to spoof. Therefore, cameras can be installed in places such as shopping malls, train stations, streets, major traffic arteries, and airports to capture images of pedestrians' walking postures, thereby extracting their gait features to determine their identity.

[0003] Existing technologies disclose a training method, a gait recognition method, and a storage device for pedestrian gait recognition in videos. This technology extracts gait features based on a single gait contour map. However, the contour map can only capture contour edge information and ignores the motion information inside the contour. It is precisely because of the lack of this important information that the recognition accuracy is greatly reduced. Furthermore, directly applying traditional convolutional neural networks to individual identity recognition methods based on gait feature information results in the output of feature maps with low semantic features, leading to low recognition accuracy. Existing gait recognition methods are not strong in resisting interference. If an individual's external morphology changes, such as wearing a coat or carrying a backpack, the recognition accuracy is not high. Summary of the Invention

[0004] The purpose of this invention is to provide a pedestrian identification method, system, and storage medium that improves the identification accuracy of gait feature image-based identification methods.

[0005] To achieve the above objectives, the present invention provides a pedestrian identification method, which includes the following steps:

[0006] Step S1: Obtain the image sequence to be recognized;

[0007] Step S2: Extract pedestrian gait contour image features and dense optical flow image features from the image sequence to be identified;

[0008] Step S3: The gait contour image features and dense optical flow image features are fused according to a first adaptive ratio to obtain a first SIL-FLOW feature map, and the gait contour image features and dense optical flow image features are fused according to a second adaptive ratio to obtain a second SIL-FLOW feature map, wherein the weighting ratio of the gait contour image features in the first adaptive ratio and the second adaptive ratio is set to a, the weighting ratio of the dense optical flow image features is set to b, and a+b=1;

[0009] Furthermore, when the pedestrian in the image is in an upright position, the gait information provided by the gait contour image features is richer than that of the dense optical flow image features, a > b;

[0010] When a pedestrian is not in an upright position, the gait information provided by dense optical flow image features is richer than that of gait contour image features, a < b;

[0011] Step S4: Input the first SIL-FLOW feature map and the second SIL-FLOW feature map into the two channels of the dual-stream feature extraction network respectively to obtain two feature vector maps. Then, fuse the two feature vector maps to obtain the fused features.

[0012] Step S5: After encoding the fused features into the gait feature encoding structure, the system outputs a feature vector corresponding to the image sequence to be identified via a fully connected neural network. The pedestrian identity is determined based on the output feature vector corresponding to the image sequence to be identified. The steps include:

[0013] The gait feature encoding structure consists of six identical gait feature encoders connected sequentially. The first gait feature encoder is connected to a two-stream feature extraction network, and the last gait feature encoder is connected to a fully connected neural network.

[0014] Each gait feature encoder includes residual branch 1, residual branch 2, residual branch 3, a multi-head attention layer, a feedforward network layer, a first summing layer, and a second summing layer.

[0015] The multi-head attention layer, the first summing layer, the feedforward network layer, and the second summing layer are connected in sequence;

[0016] The linear mapping at the input end of the multi-head attention layer generates three input heads: Q, K, and V. , , ;

[0017] One end of residual branch 1 is connected to the input of the multi-head attention layer, and the other end of residual branch 1 is connected to the first summing layer, which adds the input and output of the multi-head attention layer.

[0018] One end of residual branch 2 is connected to the input of the feedforward network layer, and the other end of residual branch 2 is connected to the second summing layer to add the input and output of the feedforward network layer.

[0019] One end of residual branch 3 is connected to the input of the multi-head attention layer, and the other end of residual branch 3 is connected to the output of the second summing layer, so that the input of the multi-head attention layer and the output of the second summing layer are fused.

[0020] A fully connected neural network consists of three sequentially connected fully connected layers, each containing 200 neurons.

[0021] The input of the first fully connected layer is connected to the output of the last gait feature encoder. The last fully connected layer outputs the feature vector corresponding to the image sequence to be identified, and the pedestrian identity is determined based on the feature vector.

[0022] Furthermore, the gait contour image features are obtained in the following way:

[0023] The pedestrian gait contour image is obtained by subtracting two adjacent images in the image sequence to be identified using the following formula.

[0024]

[0025]

[0026] In the formula, (i, j) represent the coordinates of each pixel in the image sequence to be identified. Representative moment Background, Indicates time The video frames, Indicates time The extracted binary image of the moving target. Indicates the segmentation threshold;

[0027] The pedestrian gait contour image is first subjected to morphological processing and then normalization processing to obtain pedestrian gait contour image features.

[0028] Furthermore, the dense optical flow image features are obtained in the following way:

[0029] Assuming the selected pixel brightness remains constant, the coordinates of the selected optical flow pixel in the initial image are: The position of this pixel in the terminating image is Then the optical flow vector Represented as: The optical flow vector The dense optical flow image is obtained by mapping it to the HSV color space and then visualizing it.

[0030] The dense optical flow image is then normalized to obtain dense optical flow image features.

[0031] Furthermore, the adaptive scale is determined as follows:

[0032] When the pedestrian in the image is in an upright position, the gait information provided by the gait contour image features is richer than that provided by the dense optical flow image features. The weighting ratio of the gait contour image features is set to a = 0.7. When the pedestrian in the image is not in an upright position, the gait information provided by the dense optical flow image features is richer than that provided by the gait contour image features. The weighting ratio of the gait contour image features is set to a = 0.3, and the weighting ratio of the dense optical flow image features is set to b = 0.7.

[0033] Furthermore, the dual-stream feature extraction network comprises two channels, each consisting of three sequentially connected feature extraction modules. Each feature extraction module has the same structure, as shown below:

[0034] The feature extraction module structure includes a residual branch, a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, and an upsampling layer. The first convolutional layer, the second convolutional layer, the third convolutional layer, the fourth convolutional layer, and the upsampling layer are connected in sequence. One end of the residual branch is connected to the input of the first convolutional layer, and the other end of the residual branch is connected to the output of the upsampling layer. The SIL-FLOW feature map input from the first convolutional layer is concatenated and fused with the feature vector output from the upsampling layer to obtain the fused feature.

[0035] To achieve the above objectives, the present invention also provides a pedestrian identification system, comprising:

[0036] The acquisition module is used to acquire the image sequence to be recognized;

[0037] The extraction module is used to extract gait contour image features and dense optical flow image features from the image sequence to be identified;

[0038] The first fusion module is used to fuse the extracted gait contour image features and dense optical flow image features according to a first adaptive ratio to obtain a first SIL-FLOW feature map, and to fuse the gait contour image features and dense optical flow image features according to a second adaptive ratio to obtain a second SIL-FLOW feature map.

[0039] The second fusion module is used to input the first SIL-FLOW feature map and the second SIL-FLOW feature map into the two channels of the dual-stream feature extraction network respectively to obtain two feature vector maps, and to fuse the two feature vector maps to obtain the fused features.

[0040] The recognition module first inputs the fused features into the gait feature encoding structure, obtains the corresponding output, and then inputs it into the fully connected layer to obtain the feature vector corresponding to the image sequence to be recognized. The pedestrian's identity is then determined based on the feature vector.

[0041] Finally, the present invention also provides a computer-readable storage medium storing a computer program for a pedestrian identification method, wherein when the computer program for the pedestrian identification method is processed, it implements the steps of the pedestrian identification method.

[0042] Compared with existing technologies, its advantages are as follows:

[0043] This invention combines gait contour image features and dense optical flow image features, and employs an adaptive weighting method to extract pedestrian SIL-FLOW feature maps to obtain dense optical flow images to supplement motion information in gait features. This solves the problem that using only gait contour maps can only capture contour edge information while ignoring motion information within the contour, thus improving recognition accuracy. Furthermore, it replaces traditional convolutional mapping with a gait feature encoding network.

[0044] The integral neural network transforms the fused input feature image into a low-resolution feature map with high semantic features, ensuring that the output image contains gait semantic features highly correlated with pedestrian edges, thus further improving recognition accuracy. Furthermore, the gait feature encoder of this invention introduces residual branch 3, which effectively mitigates the gradient vanishing and exploding problems inherent in traditional deep neural networks, eliminating interference from irrelevant information in the feature map, and further improving the accuracy of pedestrian identification. Attached Figure Description

[0045] Figure 1 This is a flowchart of a pedestrian identification method according to an embodiment of the present invention;

[0046] Figure 2 This is a structural block diagram of the pedestrian identification system according to an embodiment of the present invention;

[0047] Figure 3 This is a flowchart of the SIL-FLOW feature map extraction process according to an embodiment of the present invention;

[0048] Figure 4 This is a SIL-FLOW feature map according to an embodiment of the present invention;

[0049] Figure 5 This is a structural diagram of the gait feature encoder according to an embodiment of the present invention;

[0050] Figure 6 This is a structural diagram of the dual-stream feature extraction network, gait feature encoding structure, and fully connected layer in an embodiment of the present invention. Detailed Implementation

[0051] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0052] Example 1:

[0053] like Figure 1 As shown, a preferred embodiment of the pedestrian identification method of the present invention includes the following steps:

[0054] Step S1: Obtain the image sequence to be recognized;

[0055] Step S2: Extract pedestrian gait contour image features and dense optical flow image features from the image sequence to be identified;

[0056] Step S3: Fuse the gait contour image features and dense optical flow image features according to a first adaptive ratio to obtain a first SIL-FLOW feature map; fuse the gait contour image features and dense optical flow image features according to a second adaptive ratio to obtain a second SIL-FLOW feature map;

[0057] Step S4: Input the first SIL-FLOW feature map and the second SIL-FLOW feature map into the two channels of the dual-stream feature extraction network respectively to obtain two feature vector maps. Then, fuse the two feature vector maps to obtain the fused features.

[0058] Step S5: After encoding the fused features into the gait feature encoding structure, the feature vector corresponding to the image sequence to be identified is output through a fully connected neural network. The pedestrian identity is determined based on the output feature vector corresponding to the image sequence to be identified.

[0059] This embodiment combines gait contour image features and dense optical flow image features, and uses an adaptive weighting method to extract the pedestrian's SIL-FLOW feature map to obtain dense optical flow images to supplement the motion information in the gait features. This solves the problem that using only gait contour images can only capture contour edge information while ignoring motion information inside the contour, thus improving the accuracy of recognition. Furthermore, a gait feature encoding network replaces the traditional convolutional network...

[0060] The integral neural network transforms the fused input feature image into a low-resolution feature map with high semantic features, ensuring that the output image contains gait semantic features highly correlated with pedestrian edges, thus further improving recognition accuracy. Furthermore, the gait feature encoder of this invention introduces residual branch 3, which effectively mitigates the gradient vanishing and exploding problems inherent in traditional deep neural networks, eliminating interference from irrelevant information in the feature map, and further improving the accuracy of pedestrian identification.

[0061] Example 2:

[0062] like Figure 1 As shown, a pedestrian identification method according to a preferred embodiment of the present invention includes the following steps:

[0063] Step S1: Obtain the image sequence to be recognized;

[0064] In this embodiment, multiple walking videos of the target object are captured from the front by a monitoring device. Each target object includes 6 video sequences of normal clothing, 2 video sequences of carrying a backpack, and 2 video sequences of wearing a coat. The present invention selects the first 4 normal walking image sequences as training samples, and the remaining 2 normal walking image sequences, 2 backpack carrying image sequences, and 2 coat wearing image sequences as test samples. After obtaining the individual walking video sequences from the front view, the captured front view video sequences are processed by frame cutting. Only the first 10 seconds of each video sequence are selected for frame cutting, and 5 frames are cropped from each second of video. A total of 50 images are obtained for each video sequence, and a total of 500 images are obtained for 10 videos.

[0065] Step S2: Extract pedestrian gait contour image features and dense optical flow image features from the image sequence to be identified;

[0066] In this embodiment, the pedestrian gait contour image features are obtained by differentiating two adjacent frames using the following formula:

[0067]

[0068]

[0069] In the formula, (i, j) represent the coordinates of each pixel in the image sequence to be identified. Representative moment Background, Indicates time The video frames, Indicates time The extracted binary image of the moving target. The segmentation threshold is indicated; the pedestrian gait contour image is first subjected to morphological processing to eliminate missing pixels and noise pixels inside the pedestrian contour, and then normalization processing is performed to obtain individual gait contour image features in order to extract pedestrian gait contour information.

[0070] Pedestrian dense optical flow image features were obtained in the following way:

[0071] Assuming the selected pixel brightness remains constant, the coordinates of the selected optical flow pixel in the initial image are: The position of this pixel in the terminating image is Then the optical flow vector Represented as: The optical flow vector The dense optical flow image is obtained by mapping it to the HSV color space and then visualizing it. The dense optical flow image is then normalized to obtain dense optical flow image features, so as to extract pedestrian gait motion information.

[0072] Step S3: Fuse the gait contour image features and dense optical flow image features according to a first adaptive ratio to obtain a first SIL-FLOW feature map; fuse the gait contour image features and dense optical flow image features according to a second adaptive ratio to obtain a second SIL-FLOW feature map;

[0073] In this embodiment, the adaptive ratio is obtained as follows: To improve the fusion efficiency of pedestrian contour image features and dense optical flow image features, the algorithm of this invention adaptively determines the weighting ratio based on the pedestrian gait stage. When the human body is in an upright state, the gait information provided by the gait contour map is richer than that of the dense optical flow map. The weighting ratio of the gait contour map is set accordingly. The value is 0.7, which is the weighting ratio of the dense optical flow map. The value is 0.3; when the human body is in a non-upright state, the gait information provided by the dense optical flow map is richer than that of the gait profile map, and the weighting ratio of the gait profile map is set. The value is 0.3, which is the weighting ratio of the dense optical flow map. The value is 0.7.

[0074] Step S4: Input the first SIL-FLOW feature map and the second SIL-FLOW feature map into the two channels of the dual-stream feature extraction network respectively to obtain two feature vector maps. Then, fuse the two feature vector maps to obtain the fused features.

[0075] In this embodiment, the different SIL-FLOW feature maps input to the two-stream feature extraction network are obtained as follows: for the SIL-FLOW feature map input to the first-stream channel in the two-stream feature extraction network, its gait contour image feature weighting ratio is... have:

[0076] Its dense optical flow image feature weighting ratio have:

[0077]

[0078] in, This indicates that the current frame is the [number]th frame in the entire sequence. frame, This indicates that the current number is the [number]. The gait cycle. After processing the pedestrian gait sequence, the first frame of the frame sequence is in an upright position. Each subsequent gait cycle occupies 7 frames, therefore the human body is in an upright position in the 4th frame. In the entire sequence, we set the gait cycle as follows: Frame and the The human body is in an upright position in the frame, among which This indicates the current cycle number. To improve the model's robustness under drastic environmental changes, the weighted ratio of gait contour image features is calculated for the input SIL-FLOW feature map of the second stream in the two-stream feature extraction network within the same frame. Its dense optical flow image feature weighting ratio ,Right now:

[0079]

[0080]

[0081] in, This indicates that the current frame is the [number]th frame in the entire sequence. frame, This indicates that the current number is the [number]. One gait cycle.

[0082] Step S5: After encoding the fused features into the gait feature encoding structure, the feature vector corresponding to the image sequence to be identified is output through a fully connected neural network. The pedestrian identity is determined based on the output feature vector corresponding to the image sequence to be identified.

[0083] In this embodiment, the gait encoder introduces residual branch 3, which effectively mitigates the gradient vanishing and exploding problems existing in traditional deep neural networks and eliminates interference from irrelevant information in the feature map; there are three fully connected layers, each with 200 neurons, for a total of 600 neurons.

[0084] Example 3:

[0085] like Figure 2 As shown, a preferred embodiment of the present invention provides a pedestrian identification system, comprising:

[0086] The acquisition module is used to acquire the image sequence to be recognized;

[0087] The extraction module is used to extract gait contour image features and dense optical flow image features from the image sequence to be identified;

[0088] The first fusion module is used to fuse the extracted gait contour image features and dense optical flow image features according to a first adaptive ratio to obtain a first SIL-FLOW feature map, and to fuse the gait contour image features and dense optical flow image features according to a second adaptive ratio to obtain a second SIL-FLOW feature map.

[0089] The second fusion module is used to input the first SIL-FLOW feature map and the second SIL-FLOW feature map into the two channels of the dual-stream feature extraction network respectively to obtain two feature vector maps, and to fuse the two feature vector maps to obtain the fused features.

[0090] The recognition module first inputs the fused features into the gait feature encoding structure, obtains the corresponding output, and then inputs it into the fully connected layer to obtain the feature vector corresponding to the image to be recognized. The pedestrian's identity is then determined based on this feature vector.

[0091] This embodiment combines gait contour image features and dense optical flow image features, and uses an adaptive weighting method to extract the pedestrian's SIL-FLOW feature map to obtain dense optical flow images to supplement the motion information in the gait features. This solves the problem that using only gait contour images can only capture contour edge information while ignoring motion information inside the contour, thus improving the accuracy of recognition. Furthermore, a gait feature encoding network replaces the traditional convolutional network...

[0092] The integral neural network transforms the fused input feature image into a low-resolution feature map with high semantic features, ensuring that the output image contains gait semantic features highly correlated with pedestrian edges, thus further improving recognition accuracy. Furthermore, the gait feature encoder of this invention introduces residual branch 3, which effectively mitigates the gradient vanishing and exploding problems inherent in traditional deep neural networks, eliminating interference from irrelevant information in the feature map, and further improving the accuracy of pedestrian identification.

[0093] Example 4:

[0094] This embodiment also provides a preferred embodiment of a computer-readable storage medium storing a computer program for a pedestrian identification method, wherein when the computer program for the pedestrian identification method is processed, it implements the steps of the pedestrian identification method.

[0095] In summary, this invention combines gait contour image features and dense optical flow image features, and employs an adaptive weighting method to extract individual SIL-FLOW feature maps to obtain dense optical flow images to supplement motion information in gait features. This solves the problem that using only gait contour maps can only capture contour edge information while ignoring motion information within the contour, thus improving recognition accuracy. Furthermore, a gait feature encoding network replaces the traditional convolutional neural network, converting the input fused feature image into a low-resolution, high-semantic-feature feature map. This ensures the output image contains gait semantic features highly correlated with pedestrian edges, further improving recognition accuracy. Additionally, the gait feature encoder of this invention introduces residual branch 3, which effectively mitigates the gradient vanishing and exploding problems present in traditional deep neural networks, eliminating interference from irrelevant information in the feature map, further improving pedestrian identification accuracy.

[0096] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and substitutions can be made without departing from the technical principles of the present invention, and these improvements and substitutions should also be considered within the scope of protection of the present invention.

Claims

1. A method for pedestrian identification, characterized in that, The method includes the following steps: Step S1: Obtain the image sequence to be recognized; Step S2: Extract pedestrian gait contour image features and dense optical flow image features from the image sequence to be identified; Step S3: The gait contour image features and dense optical flow image features are fused according to a first adaptive ratio to obtain a first SIL-FLOW feature map, and the gait contour image features and dense optical flow image features are fused according to a second adaptive ratio to obtain a second SIL-FLOW feature map, wherein the weighting ratio of the gait contour image features in the first adaptive ratio and the second adaptive ratio is set to a, the weighting ratio of the dense optical flow image features is set to b, and a+b=1; Furthermore, when the pedestrian in the image is in an upright position, the gait information provided by the gait contour image features is richer than that of the dense optical flow image features, a > b; When a pedestrian is not in an upright position, the gait information provided by dense optical flow image features is richer than that of gait contour image features, a < b; Step S4: Input the first SIL-FLOW feature map and the second SIL-FLOW feature map into the two channels of the dual-stream feature extraction network respectively to obtain two feature vector maps. Then, fuse the two feature vector maps to obtain the fused features. Step S5: After encoding the fused features into the gait feature encoding structure, the system outputs a feature vector corresponding to the image sequence to be identified via a fully connected neural network. The pedestrian identity is determined based on the output feature vector corresponding to the image sequence to be identified. The steps include: The gait feature encoding structure consists of six identical gait feature encoders connected sequentially. The first gait feature encoder is connected to a two-stream feature extraction network, and the last gait feature encoder is connected to a fully connected neural network. Each gait feature encoder includes residual branch 1, residual branch 2, residual branch 3, a multi-head attention layer, a feedforward network layer, a first summing layer, and a second summing layer. The multi-head attention layer, the first summing layer, the feedforward network layer, and the second summing layer are connected in sequence; The linear mapping at the input end of the multi-head attention layer generates three input heads: Q, K, and V. , , ; One end of residual branch 1 is connected to the input of the multi-head attention layer, and the other end of residual branch 1 is connected to the first summing layer, which adds the input and output of the multi-head attention layer. One end of residual branch 2 is connected to the input of the feedforward network layer, and the other end of residual branch 2 is connected to the second summing layer to add the input and output of the feedforward network layer. One end of residual branch 3 is connected to the input of the multi-head attention layer, and the other end of residual branch 3 is connected to the output of the second summing layer, so that the input of the multi-head attention layer and the output of the second summing layer are fused. A fully connected neural network consists of three sequentially connected fully connected layers, each containing 200 neurons. The input of the first fully connected layer is connected to the output of the last gait feature encoder. The last fully connected layer outputs the feature vector corresponding to the image sequence to be identified, and the pedestrian identity is determined based on the feature vector.

2. The pedestrian identification method according to claim 1, characterized in that, The pedestrian gait contour image features in step S2 are obtained in the following way: The pedestrian gait contour image is obtained by subtracting two adjacent images in the image sequence to be identified using the following formula. In the formula, (i, j) represent the coordinates of each pixel in the image sequence to be identified. Representative moment Background, Indicates time The video frames, Indicates time The extracted binary image of the moving target. Indicates the segmentation threshold; The pedestrian gait contour image is first subjected to morphological processing and then normalization processing to obtain pedestrian gait contour image features.

3. The pedestrian identification method according to claim 1, characterized in that, The dense optical flow image features in step S2 are obtained as follows: Assuming the selected pixel brightness remains constant, the coordinates of the selected optical flow pixel in the initial image are: The position of this pixel in the ending image is Then the optical flow vector Represented as: The optical flow vector The dense optical flow image is obtained by mapping it to the HSV color space and then visualizing it. The dense optical flow image is then normalized to obtain dense optical flow image features.

4. The pedestrian identification method according to claim 3, characterized in that, When pedestrians are in the image When the pedestrian is in an upright position, the gait information provided by the gait contour image features is richer than that provided by the dense optical flow image features. The weighting ratio of the gait contour image features is set to a=0.7, and the weighting ratio of the dense optical flow image features is set to b=0.

3. When the pedestrian is in a non-upright position, the gait information provided by the dense optical flow image features is richer than that provided by the gait contour image features. The weighting ratio of the gait contour image features is set to a=0.3, and the weighting ratio of the dense optical flow image features is set to b=0.

7.

5. The pedestrian identification method according to claim 1, characterized in that, Step S4 The dual-stream feature extraction network consists of two channels, each composed of three sequentially connected feature extraction modules. Each feature extraction module has the same structure, as shown below: The feature extraction module structure includes a residual branch, a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, and an upsampling layer. The first convolutional layer, the second convolutional layer, the third convolutional layer, the fourth convolutional layer, and the upsampling layer are connected in sequence. One end of the residual branch is connected to the input of the first convolutional layer, and the other end of the residual branch is connected to the output of the upsampling layer. The SIL-FLOW feature map input from the first convolutional layer is concatenated and fused with the feature vector output from the upsampling layer to obtain the fused feature.

6. A pedestrian identification system, characterized in that, The system is based on the method according to any one of claims 1 to 5, and the system comprises: The acquisition module is used to acquire the image sequence to be recognized; The extraction module is used to extract gait contour image features and dense optical flow image features from the image sequence to be identified; The first fusion module is used to fuse the extracted gait contour image features and dense optical flow image features according to a first adaptive ratio to obtain a first SIL-FLOW feature map, and to fuse the gait contour image features and dense optical flow image features according to a second adaptive ratio to obtain a second SIL-FLOW feature map. The second fusion module is used to input the first SIL-FLOW feature map and the second SIL-FLOW feature map into the two channels of the dual-stream feature extraction network respectively to obtain two feature vector maps, and to fuse the two feature vector maps to obtain the fused features. The recognition module first inputs the fused features into the gait feature encoding structure, obtains the corresponding output, and then inputs it into the fully connected layer to obtain the feature vector corresponding to the image sequence to be recognized. The pedestrian's identity is then determined based on the feature vector.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the pedestrian identification method according to any one of claims 1 to 5.

Images