Occluded pedestrian re-identification method, system and device based on three-dimensional information and semantic features and medium

Abstract

The invention discloses an occluded pedestrian re-identification method, system and device based on three-dimensional information and semantic features, and a medium, and the method comprises the steps: obtaining an occluded pedestrian data set which comprises an occluded pedestrian image and corresponding pedestrian label information; performing data format conversion on the occluded pedestrian image to obtain occluded pedestrian image tensor data; an initial pedestrian re-recognition model is constructed, the initial pedestrian re-recognition model comprises a multi-dimensional feature extraction module, a semantic feature extraction module and a feature splicing module which are connected in sequence, and the feature extraction module comprises a spatial feature extraction module and a three-dimensional point cloud feature extraction module which are arranged in parallel; and training and optimizing the initial pedestrian re-identification model based on the occluded pedestrian image tensor data, and executing an occluded pedestrian re-identification task of the to-be-identified image based on the trained pedestrian re-identification model. According to the technical scheme, the pedestrian matching precision is improved, and the cross-domain adaptation problem existing in an additional model is solved.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision technology, and in particular relates to a method, system, device and medium for re-identifying occluded pedestrians based on three-dimensional information and semantic features. Background Art

[0002] Pedestrian re-identification is a classic downstream task of computer vision, which aims to match the same pedestrian across different camera perspectives and scenes. It has important applications in surveillance, security and intelligent traffic management. With the rapid development of deep learning technology, the use of deep learning to handle pedestrian re-identification tasks has achieved remarkable results. However, most traditional deep learning pedestrian re-identification methods use the entire pedestrian image as input and extract the features of the entire image as matching information. However, in real scenes, the target person is often occluded by various obstacles such as other pedestrians, vehicles or buildings, resulting in incomplete or lost pedestrian feature information. In addition, these occlusion information is accompanied by noise information, which makes it easy for the model to learn wrong matching information, resulting in a decrease in the matching accuracy of the pedestrian re-identification model. Therefore, solving the problem of occluded pedestrian re-identification has important research value and practical significance in the real scene application of pedestrian re-identification.

[0003] Existing occluded re-identification methods can be roughly divided into two types, one is based on external clues, and the other is based on part-to-part matching. The first method uses external clues to align body part features, such as pose estimation, human body parsing, instance segmentation, etc., but these methods all require the use of additional models for auxiliary recognition, which increases the recognition reasoning time, and the training data sets used by these additional models are mostly unoccluded, which has cross-domain problems with occluded pedestrian recognition and cannot be well applied to occluded recognition tasks. The second method is to extract the overall features of pedestrians, then divide them into several parts horizontally and evenly, and perform part-to-part matching, but this method does not use the semantic information of the human body, such as head-to-head matching and upper body-to-upper body matching, but roughly divides the overall features of pedestrians evenly, which will cause the system to mistakenly match different body parts, resulting in a decrease in the matching accuracy of the pedestrian re-identification model. In addition, the above two methods only use the two-dimensional feature information of pedestrians, and do not introduce the three-dimensional information in the real scene, which greatly limits the matching ability of the pedestrian re-identification model. Summary of the invention

[0004] The purpose of the present invention is to provide a method, system, device and medium for re-identifying occluded pedestrians based on three-dimensional information and semantic features, so as to solve the problems existing in the above-mentioned prior art.

[0005] To achieve the above object, the present invention provides a method for re-identifying an occluded pedestrian based on three-dimensional information and semantic features, comprising: Acquire an occluded pedestrian dataset, wherein the occluded pedestrian dataset includes occluded pedestrian images and corresponding pedestrian label information; Performing data format conversion on the occluded pedestrian image to obtain occluded pedestrian image tensor data; Constructing an initial person re-identification model; the initial person re-identification model comprises a multi-dimensional feature extraction module, a semantic feature extraction module and a feature splicing module connected in sequence, wherein the feature extraction module comprises a spatial feature extraction module and a three-dimensional point cloud feature extraction module arranged in parallel; The initial pedestrian re-identification model is trained and optimized based on the occluded pedestrian image tensor data, and the occluded pedestrian re-identification task of the image to be identified is performed based on the trained pedestrian re-identification model.

[0006] Optionally, before converting the data format of the occluded pedestrian image, the occluded pedestrian image is randomly cropped, randomly flipped and randomly erased to obtain a preprocessed occluded pedestrian image, and the data format of the preprocessed occluded pedestrian image is converted to obtain occluded pedestrian image tensor data.

[0007] Optionally, the training and optimizing the initial pedestrian re-identification model based on the occluded pedestrian image tensor data specifically includes: Inputting the occluded pedestrian image tensor data into the spatial feature extraction module to extract spatial features, obtain spatial feature maps of different scales, and perform feature fusion on the spatial feature maps of different scales to obtain a fused spatial feature map; Inputting the fused spatial feature map into the semantic feature extraction module, performing four-class classification on each pixel in the fused spatial feature map to obtain a plurality of classification feature maps, wherein the classification feature maps include a head feature map, an upper body feature map, a lower body feature map, and a background feature map; Determine the semantic category corresponding to each classification feature map, aggregate pixels of the same semantic category, generate a semantic part mask of the corresponding semantic part, aggregate each of the semantic part masks, and obtain a pedestrian foreground mask; Performing matrix multiplication operations on the pedestrian foreground mask and each of the semantic part masks and the fused spatial feature map respectively to obtain corresponding two-dimensional foreground features; Inputting the occluded pedestrian image tensor data into a three-dimensional point cloud feature extraction module to obtain three-dimensional point cloud information; Performing three-dimensional information guidance on the three-dimensional point cloud information and the two-dimensional foreground features to obtain pedestrian foreground features after three-dimensional information guidance; Using the pedestrian foreground mask and each of the semantic part masks to generate a signal for each semantic feature as to whether it is blocked, and splicing the unblocked semantic features with the pedestrian foreground features guided by the three-dimensional information to obtain pedestrian splicing features; The pedestrian label information corresponding to the pedestrian splicing features is used as a supervision signal to perform supervised training on the model, and is trained according to the target loss function to obtain a trained pedestrian re-identification model.

[0008] Optionally, the performing three-dimensional information guidance on the three-dimensional point cloud information and the two-dimensional foreground features specifically includes: The three-dimensional point cloud information is converted into three-dimensional point cloud features through a graph neural convolutional network, and the two-dimensional foreground features and the three-dimensional point cloud features are projected into the feature space of the same dimension using an MLP linear transformation layer. The distance between the three-dimensional point cloud features and the two-dimensional foreground features of the same pedestrian label information is shortened through comparative learning technology, and the distance between the three-dimensional point cloud features and the two-dimensional foreground features of different pedestrian label information is increased to obtain the pedestrian foreground features guided by the three-dimensional information.

[0009] Optionally, the target loss function includes a pedestrian classification loss function and a triplet loss function.

[0010] In addition, the present invention also provides an occluded pedestrian re-identification system based on three-dimensional information and semantic features, comprising: A data acquisition module is used to obtain a data set of occluded pedestrians, wherein the data set includes an occluded pedestrian image and corresponding pedestrian label information; perform data format conversion on the occluded pedestrian image to obtain occluded pedestrian image tensor data; A model training module is used to construct an initial pedestrian re-identification model; the initial pedestrian re-identification model includes a multi-dimensional feature extraction module, a semantic feature extraction module and a feature splicing module connected in sequence, and the feature extraction module includes a spatial feature extraction module and a three-dimensional point cloud feature extraction module set in parallel; The pedestrian re-identification module is used to train and optimize the initial pedestrian re-identification model according to the occluded pedestrian image tensor data, and perform the occluded pedestrian re-identification task of the image to be identified based on the trained pedestrian re-identification model.

[0011] An electronic device includes a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute a method for re-identifying an occluded pedestrian based on three-dimensional information and semantic features.

[0012] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements a method for re-identifying an occluded pedestrian based on three-dimensional information and semantic features.

[0013] The technical effects of the present invention are: The present invention eliminates the noise interference of obstacles by setting up a semantic feature extraction module, and can correctly match local features containing semantic information of pedestrians; through three-dimensional point cloud feature extraction and three-dimensional information guidance, two-dimensional foreground features similar to three-dimensional features are extracted, thereby introducing three-dimensional information of pedestrians into two-dimensional foreground features, and through parameter learning, the model can extract two-dimensional foreground features of pedestrians containing three-dimensional feature information without three-dimensional information guidance in the inference stage, thereby improving the accuracy of pedestrian matching. In addition, the model of the present invention is trained based on the occluded pedestrian recognition data set, thus solving the cross-domain adaptation problem existing in the additional model. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] 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 will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0015] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings: Figure 1 is an overall structural diagram of the method for re-identifying an occluded pedestrian in an embodiment of the present invention; Figure 2 is a structural diagram of a pixel classifier module in an embodiment of the present invention; Figure 3 is a structural diagram of a three-dimensional information guidance module in an embodiment of the present invention; Figure 4 This is a flowchart of re-identification of an occluded pedestrian in an embodiment of the present invention. DETAILED DESCRIPTION

[0016] Various exemplary embodiments of the present invention will now be described in detail. This detailed description should not be considered as limiting the present invention, but should be understood as a more detailed description of certain aspects, features, and embodiments of the present invention.

[0017] It should be understood that the terms described in the present invention are only for describing special embodiments and are not intended to limit the present invention. In addition, for the numerical range in the present invention, it should be understood that each intermediate value between the upper and lower limits of the scope is also specifically disclosed. Each smaller range between the intermediate value in any stated value or stated range and any other stated value or intermediate value in the described range is also included in the present invention. The upper and lower limits of these smaller ranges can be independently included or excluded in the scope.

[0018] It will be apparent to those skilled in the art that various modifications and variations may be made to the specific embodiments of the present invention description without departing from the scope or spirit of the present invention. Other embodiments derived from the present invention description will be apparent to those skilled in the art. The present application description and examples are exemplary only.

[0019] The words “include,” “including,” “have,” “contain,” etc. used in this article are open-ended terms, meaning including but not limited to.

[0020] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments. Embodiment 1

[0021] like Figure 1 - Figure 4 As shown, this embodiment provides a method for re-identifying an occluded pedestrian based on three-dimensional information and semantic features, including: obtaining an occluded pedestrian dataset, the occluded pedestrian dataset including occluded pedestrian images and corresponding pedestrian label information; performing data format conversion on the occluded pedestrian images to obtain occluded pedestrian image tensor data; constructing an initial pedestrian re-identification model; the initial pedestrian re-identification model includes a multi-dimensional feature extraction module, a semantic feature extraction module and a feature stitching module connected in sequence, the feature extraction module includes a spatial feature extraction module and a three-dimensional point cloud feature extraction module arranged in parallel; training and optimizing the initial pedestrian re-identification model based on the occluded pedestrian image tensor data, and performing the occluded pedestrian re-identification task of the image to be identified based on the trained pedestrian re-identification model.

[0022] This embodiment provides a method for re-identifying occluded pedestrians based on the fusion of multiple semantic features of human bodies guided by three-dimensional point cloud information. First, the occluded pedestrian data set is prepared and pre-processed; the pedestrian image data is input into the spatial feature extraction module to extract and fuse spatial feature maps of different scales; the pixel classifier module is then used to extract the pedestrian semantic mask from the fused spatial feature map, thereby outputting the semantic features of the pedestrian's head, upper body, lower body, and foreground; the three-dimensional information is introduced into the pedestrian's two-dimensional foreground features through the three-dimensional information guidance module; finally, the pedestrian classification loss and triplet loss are used to train the model in a supervised manner, and the model network parameters are updated through back-propagation technology. This embodiment obtains more robust pedestrian features by introducing three-dimensional information and combining different semantic features, thereby improving the recognition accuracy of the model in the task of re-identifying occluded pedestrians.

[0023] The steps of this embodiment specifically include: Step 1: Prepare a dataset of occluded pedestrians, use data augmentation techniques such as random cropping, random flipping, and random erasing to expand the diversity of the dataset, and obtain the identity information labels corresponding to the pedestrians; Step 2: Input the pedestrian image into the spatial feature extraction module. The module is composed of a SwinTransformer network pre-trained on the ImageNet dataset. After the module is processed, the pedestrian spatial feature maps at different scales are obtained. The feature maps of different scales are fused using upsampling and 1x1 convolution technology to obtain a spatial feature map with 768 channels. Step 3: Input the feature map fused in step 2 into the pixel classifier module. The module consists of three convolutional layers and is trained using pre-labeled pedestrian semantic part labels. After processing by the module, the semantic category of each pixel in the fused feature map is obtained. Pixels of the same semantic category are aggregated to generate masks of the corresponding semantic parts. The output semantic part mask is matrix-multiplied with the spatial feature map after pedestrian fusion to obtain the head features, upper body features, and lower body features corresponding to the pedestrian. The three semantic part masks are aggregated to obtain the pedestrian foreground mask, which is then multiplied with the fused spatial feature map to obtain the pedestrian foreground features.

[0024] Step 4: Input the pedestrian foreground features obtained in step 3 into the three-dimensional information guidance module. The module uses comparative learning technology to shorten the distance between the two-dimensional foreground features of the pedestrian and the three-dimensional point cloud features of the pedestrian corresponding to the same label in the projection space, and to increase the distance between the three-dimensional point cloud features of the pedestrian corresponding to the pedestrian with different labels in the projection space, thereby introducing the three-dimensional information of the pedestrian into the two-dimensional foreground features.

[0025] Step 5: Use the semantic mask obtained in step 3 to generate a signal for each semantic feature whether the feature is occluded, and splice the unoccluded semantic features with the pedestrian foreground features guided by the three-dimensional information obtained in step 4. Use the pedestrian's identity label information as a supervision signal, use the pedestrian classification loss and triplet loss to train the model in a supervised manner, and update the model parameters through back propagation technology.

[0026] This embodiment eliminates the noise interference of obstacles by setting up a semantic feature extraction module, and can correctly match local features containing semantic information of pedestrians; through three-dimensional point cloud feature extraction and three-dimensional information guidance, two-dimensional foreground features similar to three-dimensional features are extracted, thereby introducing the three-dimensional information of pedestrians into the two-dimensional foreground features, and through parameter learning, the model can extract the two-dimensional foreground features of pedestrians containing three-dimensional feature information without the guidance of three-dimensional information in the inference stage, thereby improving the accuracy of pedestrian matching. In addition, the model of this embodiment is trained based on the occluded pedestrian recognition data set, thus solving the cross-domain adaptation problem existing in the additional model.

[0027] The specific implementation process of this embodiment includes: Step 1: Prepare the occluded pedestrian dataset: The original data for pedestrian re-identification is usually surveillance video. It is necessary to use the target detection algorithm to extract the pedestrian detection frame from the surveillance video and crop it to serve as the input data of the pedestrian re-identification model. This embodiment uses Occluded_Duke and Occluded_REID, two occluded pedestrian datasets that have been cropped and annotated with pedestrian label information, as experimental data. The occluded pedestrian dataset is preprocessed and converted into tensor data through data enhancement techniques such as random cropping, random flipping, and random erasing.

[0028] Step 2: Spatial feature extraction: The pedestrian image tensor is input into the spatial feature extraction module, which is composed of a SwinTransformer network pre-trained on the ImageNet dataset. After being processed by the SwinTransformer network, spatial feature maps of different scales are obtained. The specific expression is .in The output spatial feature maps of different layers of the SwinTransformer network have 96, 192, 384, and 768 channels. The feature maps of different scales are upsampled and uniformly adjusted to a spatial feature map with a height of 64 and a width of 32. They are concatenated along the channel dimension and then input into a convolution layer with a convolution kernel size of 1 and a stride of 1 for feature fusion. Finally, the fused spatial feature map with a dimension of 768 is output. The specific expression of the fusion operation is: .

[0029] Step 3: Feature map pixel classification: The three convolutional layers of the pixel classifier module are used to classify each pixel token of the fused spatial feature map into four regions: head, upper body, lower body, and background. The labeled pedestrian semantic labels are used as supervisory signals to train the pixel classification loss. After being processed by the pixel classifier module, the semantic category corresponding to each pixel of the fused feature map is obtained. Pixels of the same semantic category are aggregated to generate masks for the corresponding semantic parts. , perform matrix multiplication operation on the output semantic part mask and the spatial feature map after pedestrian fusion to obtain the head feature corresponding to the pedestrian , upper body features , lower body features The specific expression is ,in They represent the head, upper body, and lower body parts of the pedestrian semantics respectively. The masks corresponding to the three semantic parts are max-aggregated to obtain the pedestrian foreground mask. The foreground mask and the fused spatial feature map are matrix-multiplied to obtain the pedestrian foreground feature. The specific expression is: ,in, Represents the foreground features of pedestrians.

[0030] Step 4: 3D information guidance: Generate 3D point cloud information of pedestrians through the 3D point cloud generation model, and send the 3D point cloud information and the 2D foreground features of pedestrians as input to the 3D information guidance module. The module converts the 3D point cloud information into 3D point cloud features through the graph neural convolution network. , and use the MLP linear transformation layer to transform the pedestrian foreground features With 3D point cloud features Projected to feature space of the same dimension Among them Represents the dimension of the feature space. The two-dimensional foreground features of the pedestrian are transformed into The 3D point cloud features of pedestrians corresponding to the same label The distance in the projection space is shortened, and the 3D point cloud features of pedestrians corresponding to their labels are The distance in the projection space is increased, so that the three-dimensional information of pedestrians is introduced into the foreground features. The specific expression of the contrastive learning method is: , in, Representation and The 3D features of pedestrians with the same identity labels, Indicates all The 3D features of pedestrians with different identity tags, is the cosine similarity calculation operation, Used to control the smoothness of similarity calculation.

[0031] Step 5: Concatenate pedestrian features, train classification and triplet loss: Use the semantic mask obtained by the pixel classifier module to generate a signal for each semantic feature whether the feature is occluded. (its value is 0 or 1, 0 means blocked, 1 means unblocked), the unblocked semantic features are spliced ​​with the pedestrian foreground features guided by three-dimensional information to obtain the pedestrian splicing features. The specific expression is: ,in Then, the pedestrian identity label information is used as a supervisory signal to train the model in a supervised manner, using the pedestrian classification loss With triplet loss To train and update the parameters.

[0032] The specific expression of classification loss is: , where T represents the concatenated pedestrian features, represents the cross entropy loss; The specific expression of triplet loss is: , Among them, T represents the concatenated pedestrian features, the subscript a represents a pedestrian benchmark sample, p represents a pedestrian positive sample with the same label as a, and n represents a pedestrian negative sample with a different label from a.

[0033] The global loss of the model is , Represents the contribution weight of classification loss, set to 0.5.

[0034] The model proposed in this embodiment is iteratively trained based on the above global loss formula, and all network parameters are updated.

[0035] This embodiment proposes a feature map pixel classifier for extracting semantic masks, generating local features with pedestrian semantic information, and obtaining a signal of whether the local features are occluded based on the semantic masks, aiming to eliminate the noise interference of obstacles and correctly match the local features containing pedestrian semantic information.

[0036] This embodiment proposes a three-dimensional information guidance module. During the training stage, the distance between the three-dimensional point cloud features of the pedestrian and the two-dimensional foreground features of the pedestrian in the projection space is shortened, aiming to allow the model to extract two-dimensional foreground features that are close to the three-dimensional features, thereby introducing the three-dimensional information of the pedestrian into the two-dimensional foreground features. Through parameter learning, the model can extract the two-dimensional foreground features of the pedestrian containing three-dimensional feature information without the need for three-dimensional information guidance during the inference stage, thereby improving the accuracy of pedestrian matching.

[0037] This embodiment uses the pixel classifier and contrast learning technology within the model to make the pedestrian re-identification model not need to provide additional external clues for model-assisted recognition, thus shortening unnecessary recognition reasoning time. In addition, the model is trained based on the occluded pedestrian recognition dataset, thus solving the cross-domain adaptation problem of additional models.

[0038] In practice, this embodiment provides a system for re-identifying an occluded pedestrian based on three-dimensional information and semantic features, including: A data acquisition module is used to obtain a data set of occluded pedestrians, wherein the data set includes an occluded pedestrian image and corresponding pedestrian label information; perform data format conversion on the occluded pedestrian image to obtain occluded pedestrian image tensor data; A model training module is used to construct an initial pedestrian re-identification model; the initial pedestrian re-identification model includes a multi-dimensional feature extraction module, a semantic feature extraction module and a feature splicing module connected in sequence, and the feature extraction module includes a spatial feature extraction module and a three-dimensional point cloud feature extraction module set in parallel; The pedestrian re-identification module is used to train and optimize the initial pedestrian re-identification model according to the occluded pedestrian image tensor data, and perform the occluded pedestrian re-identification task of the image to be identified based on the trained pedestrian re-identification model.

[0039] In practice, this embodiment provides an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute the method for re-identifying occluded pedestrians based on three-dimensional information and semantic features.

[0040] In practice, a computer-readable storage medium is provided in this embodiment, which stores a computer program. When the computer program is executed by a processor, the method for re-identifying an occluded pedestrian based on three-dimensional information and semantic features is implemented.

[0041] The above is only a preferred specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for re-identifying occluded pedestrians based on three-dimensional information and semantic features, characterized in that: include: Acquire an occluded pedestrian data set, wherein the occluded pedestrian data set includes occluded pedestrian images and corresponding pedestrian label information; Performing data format conversion on the occluded pedestrian image to obtain occluded pedestrian image tensor data; Constructing an initial person re-identification model; the initial person re-identification model comprises a multi-dimensional feature extraction module, a semantic feature extraction module and a feature splicing module connected in sequence, wherein the feature extraction module comprises a spatial feature extraction module and a three-dimensional point cloud feature extraction module arranged in parallel; The initial pedestrian re-identification model is trained and optimized based on the occluded pedestrian image tensor data, and the occluded pedestrian re-identification task of the image to be identified is performed based on the trained pedestrian re-identification model.

2. The method for re-identifying an occluded pedestrian based on three-dimensional information and semantic features according to claim 1, characterized in that: Before converting the data format of the occluded pedestrian image, the method also includes randomly cropping, randomly flipping and randomly erasing the occluded pedestrian image to obtain a preprocessed occluded pedestrian image, and converting the data format of the preprocessed occluded pedestrian image to obtain occluded pedestrian image tensor data.

3. The method for re-identifying an occluded pedestrian based on three-dimensional information and semantic features according to claim 1, characterized in that: The training and optimizing of the initial pedestrian re-identification model based on the occluded pedestrian image tensor data specifically includes: Inputting the occluded pedestrian image tensor data into the spatial feature extraction module to extract spatial features, obtain spatial feature maps of different scales, and perform feature fusion on the spatial feature maps of different scales to obtain a fused spatial feature map; Inputting the fused spatial feature map into the semantic feature extraction module, performing four-class classification on each pixel in the fused spatial feature map to obtain a plurality of classification feature maps, wherein the classification feature maps include a head feature map, an upper body feature map, a lower body feature map, and a background feature map; Determine the semantic category corresponding to each classification feature map, aggregate pixels of the same semantic category, generate a semantic part mask of the corresponding semantic part, aggregate each of the semantic part masks, and obtain a pedestrian foreground mask; Performing matrix multiplication operations on the pedestrian foreground mask and each of the semantic part masks and the fused spatial feature map respectively to obtain corresponding two-dimensional foreground features; Inputting the occluded pedestrian image tensor data into a three-dimensional point cloud feature extraction module to obtain three-dimensional point cloud information; Performing three-dimensional information guidance on the three-dimensional point cloud information and the two-dimensional foreground features to obtain pedestrian foreground features after three-dimensional information guidance; Using the pedestrian foreground mask and each of the semantic part masks to generate a signal for each semantic feature as to whether it is blocked, and splicing the unblocked semantic features with the pedestrian foreground features guided by the three-dimensional information to obtain pedestrian splicing features; The pedestrian label information corresponding to the pedestrian splicing features is used as a supervision signal to perform supervised training on the model, and is trained according to the target loss function to obtain a trained pedestrian re-identification model.

4. The method for re-identifying an occluded pedestrian based on three-dimensional information and semantic features according to claim 3, characterized in that: The providing three-dimensional information guidance for the three-dimensional point cloud information and the two-dimensional foreground features specifically includes: The three-dimensional point cloud information is converted into three-dimensional point cloud features through a graph neural convolutional network, and the two-dimensional foreground features and the three-dimensional point cloud features are projected into the feature space of the same dimension using an MLP linear transformation layer. The distance between the three-dimensional point cloud features and the two-dimensional foreground features of the same pedestrian label information is shortened through comparative learning technology, and the distance between the three-dimensional point cloud features and the two-dimensional foreground features of different pedestrian label information is increased to obtain the pedestrian foreground features guided by the three-dimensional information.

5. The method for re-identifying an occluded pedestrian based on three-dimensional information and semantic features according to claim 3, characterized in that: The target loss function includes a pedestrian classification loss function and a triplet loss function.

6. An occluded pedestrian re-identification system based on three-dimensional information and semantic features, characterized in that: include: A data acquisition module is used to obtain a data set of occluded pedestrians, wherein the data set includes images of occluded pedestrians and corresponding pedestrian label information; Performing data format conversion on the occluded pedestrian image to obtain occluded pedestrian image tensor data; A model training module is used to construct an initial pedestrian re-identification model; the initial pedestrian re-identification model includes a multi-dimensional feature extraction module, a semantic feature extraction module and a feature splicing module connected in sequence, and the feature extraction module includes a spatial feature extraction module and a three-dimensional point cloud feature extraction module set in parallel; The pedestrian re-identification module is used to train and optimize the initial pedestrian re-identification model according to the occluded pedestrian image tensor data, and perform the occluded pedestrian re-identification task of the image to be identified based on the trained pedestrian re-identification model.

7. An electronic device, characterized in that: It includes a memory and a processor, the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute a method for re-identifying an occluded pedestrian based on three-dimensional information and semantic features according to any one of claims 1-5.

8. A computer-readable storage medium, characterized in that: The computer program is stored therein, and when the computer program is executed by a processor, a method for re-identifying an occluded pedestrian based on three-dimensional information and semantic features as described in any one of claims 1 to 5 is implemented.

Images