A video pedestrian re-identification method, system and related device

Abstract

The application provides a video pedestrian re-identification method, comprising: acquiring a to-be-detected video containing a target pedestrian, and identifying different posture images of the target pedestrian in the to-be-detected video; determining image quality of each posture image, and determining an optimal posture image with the highest image quality; performing feature extraction and human contour line feature extraction on the optimal posture image in time sequence respectively, and obtaining representation features and human contour line features under the same posture respectively; fusing the representation features and the human contour line features in a channel dimension to obtain fusion features of the target pedestrian; and performing similarity calculation according to the fusion features and pedestrian features of the target pedestrian, and identifying the target pedestrian according to the calculation results from high to low similarity. The application can significantly improve the pedestrian re-identification accuracy after the posture change and the clothes change of the pedestrian. The application also provides a video pedestrian re-identification system, a computer readable storage medium and an electronic device, which have the above beneficial effects.

Description

Technical Field

[0001] This application relates to the field of image recognition, and in particular to a video pedestrian re-identification method, a video pedestrian re-identification system, and related devices. Background Technology

[0002] In existing video pedestrian recognition processes, pedestrians often change their clothes midway or due to lighting conditions, causing significant changes in human physical characteristics. This leads to a substantial decrease in the retrieval accuracy of video pedestrian re-identification, severely impacting the effectiveness of video pedestrian re-identification.

[0003] In one pedestrian re-identification scheme, the pedestrians are classified according to different shooting angles. A predetermined number of video frames with the largest unobstructed detection box area in each category are selected as the video frames to be processed. A multi-angle image feature matrix of the pedestrian to be identified is established. The features of the same region of the video frames to be processed from different angles are fused, and the matching value between the target feature matrix and the feature matrix to be identified is calculated to complete the pedestrian re-identification. However, this scheme only uses the same human body region. When encountering situations where there are significant differences in the representation information, such as when the pedestrian changes clothes, the accuracy of pedestrian re-identification retrieval under multiple angles will drop significantly, making it difficult to re-identify pedestrian targets who have changed clothes under multiple angles. Summary of the Invention

[0004] The purpose of this application is to provide a video pedestrian re-identification method, a video pedestrian re-identification system, a computer-readable storage medium, and an electronic device. By fusing cross-domain joint features obtained by multiplying pedestrian representation features of different postures and human body contour features of corresponding postures, and re-identifying pedestrians based on the fused features, the accuracy of pedestrian re-identification can be significantly improved.

[0005] To address the aforementioned technical problems, this application provides a video pedestrian re-identification method, the specific technical solution of which is as follows:

[0006] Acquire a video containing a target pedestrian and identify different body posture images of the target pedestrian in the video;

[0007] Determine the image quality of each of the aforementioned posture images, and determine the optimal posture image with the highest image quality;

[0008] The optimal posture image is subjected to feature extraction and human body contour feature extraction in chronological order to obtain the representation features and human body contour features of the target pedestrian under the same posture.

[0009] The representation features and human body contour features corresponding to different body postures are fused position by position in the channel dimension to obtain the fused features of the target pedestrian.

[0010] The similarity is calculated based on the fused features and the pedestrian features of the target pedestrian, and the target pedestrian is identified according to the calculation results from high to low similarity.

[0011] Optionally, identifying different posture images of the target pedestrian in the video to be detected includes:

[0012] The target detection algorithm and the tracking algorithm are used to determine the tracking box corresponding to the target in the video to be detected;

[0013] The pose classification model is used to classify the pose of the human target in the tracking frame to obtain different pose images of the target pedestrian.

[0014] Optionally, determining the image quality of each of the posture images includes:

[0015] Calculate the type score and position score of each of the aforementioned posture images;

[0016] A quality score is determined by multiplying the type score and the position score; the quality score is used to characterize the image quality of the posture image.

[0017] Optionally, calculating the type and location scores of the posture image includes:

[0018] The type score is determined based on the similarity between the type of the detected target in the first frame and the type of the tracked target during the tracking process.

[0019] The position is determined by the quotient of the intersection ratio of the tracking frame and the image and the distance from the tracking frame to the center point of the image.

[0020] Optionally, feature extraction and human contour feature extraction are performed on the optimal posture image in chronological order to obtain the representation features and human contour features of the target pedestrian in the same posture, including:

[0021] The optimal posture image is input into the human body contour feature extraction module in chronological order to obtain human body contour features; wherein, the human body contour feature extraction module is used to upsample the optimal posture image, and obtain a human body contour binarized image through dimensionality reduction and activation function, and then pass the human body contour binarized image and the optimal posture image through a convolutional layer, a pooling layer and an activation layer in sequence to obtain the human body contour features;

[0022] The optimal posture image is input into a feature extraction network with a residual network as its backbone in chronological order to obtain the representation features.

[0023] Optionally, after obtaining the characterization features and human contour features of the target pedestrian in the same posture, the method further includes:

[0024] The representational features and human contour features under the same posture are multiplied and fused in the channel dimension to obtain the joint features of the target pedestrian under various postures.

[0025] Optionally, the representational features and human contour features corresponding to different body postures are fused position-by-position along the channel dimension to obtain the fused features of the target pedestrian, including:

[0026] The combined features of the target pedestrian under each posture are multiplied positionally along the channel dimension to obtain the fused features of the target pedestrian.

[0027] This application also provides a video pedestrian re-identification system, including:

[0028] The acquisition module is used to acquire a video to be detected containing a target pedestrian and to identify different posture images of the target pedestrian in the video to be detected.

[0029] The image quality calculation module is used to determine the image quality of each of the posture images and to determine the optimal posture image with the highest image quality.

[0030] The feature extraction module is used to perform feature extraction and human contour feature extraction on the optimal posture image in chronological order, so as to obtain the representation features and human contour features of the target pedestrian under the same posture.

[0031] The feature fusion module is used to fuse the representation features and human body contour features corresponding to different body postures in the channel dimension to obtain the fused features of the target pedestrian.

[0032] The feature comparison module is used to calculate the similarity between the fused features and the pedestrian features of the target pedestrian, and to identify the target pedestrian according to the calculation results from high to low similarity.

[0033] This application also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method described above when it invokes the computer program in the memory.

[0034] This application provides a video pedestrian re-identification method, comprising: acquiring a video to be detected containing a target pedestrian, and identifying different posture images of the target pedestrian in the video to be detected; determining the image quality of each posture image, and determining the optimal posture image with the highest image quality; performing feature extraction and human contour feature extraction on the optimal posture image in chronological order to obtain the representation features and human contour features of the target pedestrian under the same posture; fusing the representation features and human contour features corresponding to different postures in a channel dimension to obtain the fused features of the target pedestrian; calculating the similarity based on the fused features and the pedestrian features of the target pedestrian, and identifying the target pedestrian according to the calculation results from high to low similarity.

[0035] After acquiring the video to be detected, this application identifies different posture images of the target pedestrian and selects the best posture image with the highest image quality for feature extraction and human contour feature extraction. Then, the feature extraction and human contour feature are fused to obtain a fused feature containing multiple posture features and corresponding human contour features. This feature includes representational information of the pedestrian's posture at different angles and human contour information under the corresponding postures. This feature exhibits good robustness in both the RGB color gamut corresponding to the human features and the black-and-white binary domain corresponding to the human contour features. Using this fused feature for subsequent feature comparison and retrieval can significantly improve the accuracy of pedestrian re-identification after changes in pedestrian posture and clothing.

[0036] This application also provides a video pedestrian re-identification system, a computer-readable storage medium, and an electronic device, which have the above-mentioned beneficial effects, and will not be elaborated here. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0038] Figure 1 A flowchart illustrating a video pedestrian re-identification method provided in this application embodiment;

[0039] Figure 2 This is a schematic diagram of the human body contour feature extraction module provided in an embodiment of this application;

[0040] Figure 3 This is a schematic diagram of the structure of a video pedestrian re-identification system provided in an embodiment of this application. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0042] Please refer to Figure 1 , Figure 1 A flowchart of a video pedestrian re-identification method provided in this application embodiment, the method including:

[0043] S101: Acquire a video to be detected containing the target pedestrian, and identify different body posture images of the target pedestrian in the video to be detected;

[0044] This step aims to acquire the video to be detected and identify different posture images of the target pedestrian. There are no restrictions on how the video to be detected is acquired; it can originate from a video recording device, such as a video stream captured by a camera.

[0045] Next, it is necessary to detect and track target pedestrians in the video to be detected. The target pedestrian refers to any pedestrian that needs to be identified; that is, the target pedestrian can be a pre-defined detection target or a pedestrian target existing in the video to be detected. There can be at least one target pedestrian or multiple target pedestrians.

[0046] There are no restrictions on how to identify different body postures of the target pedestrian in the video to be detected; a posture classification model can be used for posture detection.

[0047] A feasible method for identifying different poses of a target pedestrian in a video to be detected may include the following steps:

[0048] The first step is to use an object detection algorithm to determine the location and type of the target in the video to be detected;

[0049] The second step is to use an object detection algorithm to determine the detection box and use a tracking algorithm to predict the tracking box.

[0050] The third step is to use a posture classification model to classify the extracted target postures, thereby obtaining different posture images of the target pedestrian.

[0051] Object detection and tracking algorithms are used to acquire corresponding images from videos, while pose classification models are used to classify the poses of the acquired images. The object detection algorithm can employ algorithms such as YOLO to determine the type and location of each object in the video to be detected. It's important to note that the object can be any object in the video, including pedestrians and objects, and pedestrian targets are prioritized after the target type is determined. Simultaneously, starting from the first frame of the video, detection boxes are converted into tracking boxes; that is, the detection boxes and tracking boxes are in the same position in the first frame, and subsequent frames begin predicting tracking boxes. Furthermore, in subsequent frames, human targets within the tracking boxes are classified according to their poses. The pose types included in the pose classification model are not limited here; they can include frontal, side, and back views, and the pose samples can be further refined into more categories as needed. It's easy to understand that the richer the variety of pose types, the higher the detection accuracy.

[0052] Detection boxes are typically determined by a detection model, while tracking boxes are predicted by a tracking algorithm. A tracking box can be understood as representing a trajectory of the same target within the entire video. No specific restrictions are placed on the detection model or tracking algorithm used here. The goal of detection and tracking is to obtain images of the same target at different times within the video segment.

[0053] Furthermore, to ensure accuracy in pedestrian tracking during the application of bounding boxes, the Kalman algorithm can be used to predict the bounding boxes, and the Hungarian algorithm can be used to refine them through matching. The Hungarian algorithm can be applied to the optimization of bounding box tracking for multiple targets, achieved by finding the maximum matching in a bipartite graph.

[0054] S102: Determine the image quality of each of the posture images, and determine the optimal posture image with the highest image quality;

[0055] This step aims to determine the image quality of the pose images. By selecting the optimal pose image with the highest image quality for feature extraction, the accuracy of the extracted features can be improved. It should also be noted that this optimal pose image corresponds to each pose type; that is, each pose corresponds to at least one optimal pose image. This optimal pose image can be one or several high-quality images for a given pose.

[0056] In one feasible implementation method of this step, image quality can be determined by determining the type score and the location score. Specifically, the type score and location score of each pose image can be calculated, and then the quality score can be determined based on the product of the type score and the location score. This quality score is used to characterize the image quality of the pose image.

[0057] Specifically, a type score can be determined based on the similarity between the type of the detected target in the first frame and the type of the tracked target during the tracking process. The type score is used to evaluate the accuracy of detection and tracking. Specifically, it can be calculated by comparing the type of the target in the first frame with the type of the tracked target during the tracking process (which can be set to 25 frames). The similarity is obtained through feature extraction models for different body posture types. If the similarity is less than 0.7, the type score is set to 0; if the similarity is greater than 0.7, the type score is set to 1.

[0058] The position score is used to evaluate the actual quality of image capture during detection and tracking. Specifically, it is determined by the intersection-over-union ratio of the tracking bounding box and the image, and the quotient of the distance from the tracking bounding box to the center point of the image. That is, the closer the detection and tracking bounding box is to the center of the image, the larger its area, and the higher its position score.

[0059] The image quality can then be determined by the product of the location score and the type score.

[0060] Of course, those skilled in the art can use other methods to evaluate image quality, which will not be listed here, and should also be within the scope of protection of this application.

[0061] S103: Perform feature extraction and human contour feature extraction on the optimal posture image in chronological order to obtain the representation features and human contour features under the same posture.

[0062] This step aims at feature extraction, specifically extracting representational features and human body contour features. Representational features refer to the physical characteristics of the target pedestrian, but are easily affected by factors such as clothing changes. Human body contour features, on the other hand, are less affected by clothing changes.

[0063] Specifically, when performing this step, a feature extraction module and a human body contour feature extraction module can be used respectively to obtain representational features and human body contour features.

[0064] The optimal body posture image can be input into the human body contour feature extraction module in chronological order to obtain human body contour features.

[0065] The human body contour feature extraction module is used to upsample the optimal posture image and obtain a human body contour binarized image through dimensionality reduction and activation functions. The human body contour binarized image and the optimal posture image are then sequentially passed through convolutional layers and pooling layers to obtain human body contour features. Upsampling refers to image upsampling, used to enlarge the image using interpolation.

[0066] Meanwhile, the optimal posture image can be input into the feature extraction model with residual network as the backbone in chronological order to obtain the representation features.

[0067] It should be noted that the processes of extracting representational features and human contour features are independent of each other and do not interfere with each other. That is, there is no specific limitation on the execution order of the two processes. In practical applications, they can be executed simultaneously to improve the recognition efficiency when applied in this embodiment.

[0068] S104: The representation features and human body contour features corresponding to different body postures are fused positionally in the channel dimension to obtain the fused features of the target pedestrian;

[0069] This step requires feature fusion between representational features and human contour features. Since representational features and human contour features have the same dimension, fusion can be performed along the channel dimension. It can be performed by bitwise multiplication or bitwise addition; the specific fusion method can be set by those skilled in the art based on actual recognition accuracy requirements. The resulting fused features not only focus on representational information but also on contour features that are less affected by representational changes. This significantly improves retrieval accuracy in video pedestrian re-identification caused by pedestrian clothing changes.

[0070] It should be noted that the fusion features obtained in this step are the fusion features corresponding to various body postures.

[0071] S105: Calculate the similarity between the fused features and the pedestrian features of the target pedestrian, and identify the target pedestrian according to the calculation results from high to low similarity.

[0072] This step requires pedestrian re-identification based on the fused features. Specifically, it involves calculating the similarity between the pedestrian features and the target pedestrian's features. For example, this can be done by calculating the cosine distance or Euclidean distance between features to obtain feature similarity. The results are then ranked according to the similarity scores to identify potential targets and complete pedestrian re-identification. These pedestrian features are known; for instance, if a specific target pedestrian needs to be detected in a video, the known features of that target pedestrian can be compared with the fused features of each pedestrian target in the video to calculate the similarity between each pedestrian target and the target pedestrian, thus determining which targets might be that target pedestrian.

[0073] As a preferred execution method, the representational features and human contour features under the same posture can be multiplied and fused bitwise along the channel dimension to obtain the joint features of the target pedestrian under various postures. Then, the joint features of the target pedestrian under various postures are multiplied bitwise along the channel dimension to obtain the fused features of the target pedestrian. Each posture has corresponding joint features used to represent the feature information of the target pedestrian under that posture. The fused feature, on the other hand, integrates all the joint features of the target pedestrian under different postures. It can be seen that the fused feature not only contains the representational features of each posture, but also contains the human contour features corresponding to each posture. By performing pedestrian re-identification through the fused feature, the misidentification of the target caused by clothing changes can be effectively filtered out in subsequent feature comparison, thereby improving the retrieval accuracy of pedestrian re-identification.

[0074] This application embodiment acquires the video to be detected, identifies different posture images of the target pedestrian, and selects the best posture image with the highest image quality for characterization features and human contour features. Then, the characterization features and human contour features are fused to obtain a fused feature containing multiple posture features and corresponding human contour features. This feature includes characterization information of the pedestrian's posture at different angles and human contour information under the corresponding postures. This feature exhibits good robustness in both the RGB color gamut corresponding to human features and the black-and-white binary domain corresponding to human contour features. Using this fused feature for subsequent feature comparison and retrieval can significantly improve the accuracy of pedestrian re-identification after changes in pedestrian posture and clothing.

[0075] See Figure 2 , Figure 2 This is a schematic diagram of the human body contour feature extraction module provided in an embodiment of this application. Figure 2 In the flowchart of the human contour feature extraction module, Blocks 1-4 are the conv stages (convolutional layers) of the ResNet network model, which extract features through convolutional downsampling. Block 5 aims to obtain a binary image of the human contour, which can be composed of convolution, activation functions (sigmoid), pooling, etc. The transformed binary image of the human contour is multiplied bitwise with the shallow human features extracted by Blocks 1-4, and then convolved by Block 6 to obtain the final human contour features. Block 6 is mainly used to change the feature dimension. Block 7 is also a conv stage of the ResNet network model, consistent with Blocks 1-4. This branch performs deep feature extraction on the input image to obtain representation features. The deep representation features are multiplied bitwise with the contour features output by Block 6 to obtain the fused representation and contour features. The shallow features have richer texture and edges, while the deep features are the semantic features of the image.

[0076] The following describes a video pedestrian re-identification system provided by an embodiment of this application. The video pedestrian re-identification system described below can be referred to in correspondence with the video pedestrian re-identification method described above.

[0077] See Figure 2 This application also provides a video pedestrian re-identification system, characterized in that it includes:

[0078] The acquisition module is used to acquire a video to be detected containing a target pedestrian and to identify different posture images of the target pedestrian in the video to be detected.

[0079] The image quality calculation module is used to determine the image quality of each of the posture images and to determine the optimal posture image with the highest image quality.

[0080] The feature extraction module is used to perform feature extraction and human contour feature extraction on the optimal posture image in chronological order, so as to obtain the representation features and human contour features under the same posture.

[0081] The feature fusion module is used to fuse the representation features and human body contour features corresponding to different body postures in the channel dimension to obtain the fused features of the target pedestrian.

[0082] The feature comparison module is used to calculate the similarity between the fused features and the pedestrian features of the target pedestrian, and to identify the target pedestrian according to the calculation results from high to low similarity.

[0083] Based on the above embodiments, as a preferred embodiment, the acquisition module includes:

[0084] The posture recognition unit is used to determine the tracking box corresponding to the target in the video to be detected using target detection algorithm and tracking algorithm; and to classify the posture of the human target in the tracking box using a posture classification model to obtain different posture images of the target pedestrian.

[0085] Based on the above embodiments, as a preferred embodiment, the image quality calculation module includes:

[0086] A score calculation unit is used to calculate the type score and position score of each of the posture images;

[0087] An image quality calculation unit is used to determine a quality score based on the product of the type score and the position score; the quality score is used to characterize the image quality of the posture image.

[0088] Based on the above embodiments, as a preferred embodiment, the fraction calculation unit includes:

[0089] The type score calculation subunit is used to determine the type score based on the similarity between the type of the detected target in the first frame and the type of the tracked target during the tracking process.

[0090] The position score calculation subunit is used to determine the position score based on the intersection-over-interference ratio of the tracking frame and the screen, and the quotient of the distance from the tracking frame to the center point of the screen.

[0091] Based on the above embodiments, as a preferred embodiment, the feature extraction module includes:

[0092] The first extraction unit is used to input the optimal posture image into the human body contour feature extraction module in chronological order to obtain human body contour features; wherein, the human body contour feature extraction module is used to upsample the optimal posture image, and obtain a human body contour binarized image through dimensionality reduction and activation function, and then pass the human body contour binarized image and the optimal posture image through a convolutional layer, a pooling layer and an activation layer in sequence to obtain the human body contour features;

[0093] The second extraction unit is used to input the optimal posture image into a feature extraction network with a residual network as its backbone in chronological order to obtain the representation features.

[0094] Based on the above embodiments, as a preferred embodiment, it further includes:

[0095] The joint feature fusion module is used to multiply and fuse the representation features and the human body contour features under the same posture in the channel dimension to obtain the joint features of the target pedestrian under various postures.

[0096] Based on the above embodiments, as a preferred embodiment, the feature fusion module is a module used to multiply the joint features of the target pedestrian under each posture in the channel dimension to obtain the fused features of the target pedestrian.

[0097] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed, can perform the steps provided in the above embodiments. The storage medium may include various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0098] This application also provides an electronic device that may include a memory and a processor. The memory stores a computer program, and when the processor calls the computer program in the memory, it can implement the steps provided in the above embodiments. Of course, the electronic device may also include various network interfaces, power supplies, and other components.

[0099] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. As the system provided in the embodiments corresponds to the method provided in the embodiments, the description is relatively simple; relevant parts can be found in the method section.

[0100] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this application. It should be noted that those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of the claims of this application.

[0101] It should also be noted that, in this specification, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

Claims

1. A video person re-identification method, characterized in that, The method comprises the following steps: acquiring a to-be-detected video containing a target pedestrian, and identifying different posture images of the target pedestrian in the to-be-detected video; determining the image quality of each of the posture images, and determining an optimal posture image with the highest image quality; performing feature extraction and human contour line feature extraction on the optimal posture image in time sequence respectively, and obtaining representation features and human contour line features of the target pedestrian in the same posture respectively; fusing the corresponding representation features and human contour line features in different postures in the channel dimension by bit, and obtaining the fusion features of the target pedestrian; performing similarity calculation according to the fusion features and pedestrian features of the target pedestrian, and identifying the target pedestrian according to the calculation results from high to low similarity; wherein, determining the image quality of each of the posture images comprises: calculating the type score and the position score of each of the posture images; determining the quality score according to the product of the type score and the position score; the quality score is used to represent the image quality of the posture image; wherein, calculating the type score and the position score of the posture image comprises: determining the type score according to the similarity between the type of the first frame of the detection target and the type of the tracking target in the tracking process; determining the position score according to the quotient of the tracking frame and the picture intersection ratio, and the distance from the tracking frame to the center point of the picture.

2. The video person re-identification method of claim 1, wherein, Identifying different posture images of the target pedestrian in the to-be-detected video comprises: determining the tracking frame corresponding to the target in the to-be-detected video by using a target detection algorithm and a tracking algorithm; classifying the posture of the human target in the tracking frame by using a posture classification model, and obtaining different posture images of the target pedestrian. 3.The video person re-identification method of claim 1, wherein, Performing feature extraction and human contour line feature extraction on the optimal posture image in time sequence respectively, and obtaining representation features and human contour line features of the target pedestrian in the same posture respectively comprises: inputting the optimal posture image into a human contour line feature extraction module in time sequence to obtain human contour line features; wherein, the human contour line feature extraction module is used to up-sample the optimal posture image, and obtain human contour line binary images through dimension reduction and activation function, and then the human contour line binary images and the optimal posture image are sequentially input into a convolution layer, a pooling layer and an activation layer to obtain the human contour line features; inputting the optimal posture image into a feature extraction network with a residual network as the backbone in time sequence to obtain representation features.

4. The video person re-identification method of claim 1, wherein, After obtaining the representation features and human contour line features of the target pedestrian in the same posture respectively, it further comprises: multiplying and fusing the representation features and the human contour line features in the same posture in the channel dimension by bit to obtain joint features of the target pedestrian in each posture.

5. The video person re-identification method of claim 4, wherein, Fusing the corresponding representation features and human contour line features in the channel dimension by bit to obtain the fusion features of the target pedestrian comprises: multiplying the joint features of the target pedestrian in each posture in the channel dimension by bit to obtain the fusion features of the target pedestrian. 6.A video person re-identification system, characterized in that, The method comprises the following steps: an acquisition module, configured to acquire a to-be-detected video containing a target pedestrian, and identify different posture images of the target pedestrian in the to-be-detected video; An image quality calculation module is configured to determine image quality of each of the posture images and determine an optimal posture image with the highest image quality; A feature extraction module is configured to perform feature extraction and human contour line feature extraction on the optimal posture image in time sequence respectively, and obtain representation features and human contour line features of the target pedestrian in the same posture respectively; A feature fusion module is configured to fuse the representation features and the human contour line features corresponding to different postures in channel dimension by bit, and obtain fusion features of the target pedestrian; A feature comparison module is configured to perform similarity calculation according to the fusion features and pedestrian features of the target pedestrian, and identify the target pedestrian according to the calculation results from high to low similarity. The image quality calculation module comprises: A score calculation unit is configured to calculate a type score and a position score of each of the posture images; An image quality calculation unit is configured to determine a quality score according to the product of the type score and the position score; the quality score is used to represent the image quality of the posture image. The score calculation unit comprises: A type score calculation subunit is configured to determine a type score according to the similarity between the type of the first frame of the detection target and the type of the tracking target in the tracking process; A position score calculation subunit is configured to determine a position score according to the quotient of the tracking frame and the picture intersection ratio, and the distance from the tracking frame to the center point of the picture.

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

8. An electronic device, comprising: The computer program is executed by the processor to realize the steps of the video pedestrian re-identification method according to any one of claims 1-5. The computer program is executed by the processor to realize the steps of the video pedestrian re-identification method according to any one of claims 1-5.

Images