A Small Group Person Re-identification Method and System Based on Bidirectional Feature Propagation

Abstract

The present invention discloses a method and system for re-identifying a small group of people based on bidirectional feature propagation. The method includes: constructing a semantic segmentation map from a group image and performing repair and generation processing to obtain a sample image; constructing an information model to perform feature extraction processing on the sample image to obtain image feature information; performing fusion processing on the image feature information based on a bidirectional feature propagation module to obtain an image identity feature; and performing re-identification of a small group of people based on the image identity feature to obtain an identification result. The system includes: a construction module, an extraction module, a fusion module, and an identification module. By using the present invention, it is possible to improve the accuracy of re-identifying a small group of people by expanding the sample size of the group image and extracting the image identity feature of the group image. As a method and system for re-identifying a small group of people based on bidirectional feature propagation, the present invention can be widely applied to the technical field of pedestrian re-identification.

Description

Technical Field

[0001] The present invention relates to the technical field of pedestrian re-identification, and particularly to a method and system for small group pedestrian re-identification based on bidirectional feature propagation. Background Art

[0002] Small group pedestrian re-identification is a technology that takes visual information such as images and videos as input, analyzes through computer vision technology, and determines whether a specific target group exists in the target set. Its purpose is to retrieve the same group of people appearing in different surveillance cameras, so as to provide a basis for applications such as crowd behavior analysis, single-shot or cross-shot trajectory characterization, abnormal crowd detection, public area person search, and criminal investigation systems. Recently, this field has attracted more and more research attention. Compared with the pedestrian re-identification technology focusing on single individuals, small group pedestrian re-identification can utilize the mutual relationship between people, can improve the influence brought by adverse situations such as blur, occlusion, and different clothing information, and enhance the robustness and accuracy of the system. In addition, a good small group pedestrian re-identification system can also enhance the robustness of the single-person pedestrian re-identification system. However, there are two main problems in small group pedestrian re-identification technology. One is that the scale of the dataset for small group pedestrian re-identification is still small. If the scale of the dataset is small, it will severely limit the generalization and robustness of the model. The other is that the accuracy of some existing solutions is not high. The existing technology proposes a central rectangular ring ratio occurrence descriptor and a block-based ratio occurrence descriptor, models the overall distribution structure in the image, uses the covariance descriptor for the overall statistical generalization of group images, and calculates the best group match through appearance matching from this descriptor or a multi-order matching process based on multi-granularity representation and importance weights. However, the final recognition accuracy of these solutions is relatively low, mainly due to insufficient utilization of spatial layout information. Summary of the Invention

[0003] In order to solve the above technical problems, the purpose of the present invention is to provide a method and system for small group pedestrian re-identification based on bidirectional feature propagation, which can improve the accuracy of small group pedestrian re-identification by expanding the sample size of group images and extracting the image identity features of group images.

[0004] The first technical solution adopted by the present invention is: A method for small group pedestrian re-identification based on bidirectional feature propagation, comprising the following steps:

[0005] Construct a semantic segmentation map based on the group image and perform repair and generation processing to obtain a sample image;

[0006] Construct an information model to perform feature extraction processing on the sample image to obtain image feature information;

[0007] Based on the bidirectional feature propagation module, perform fusion processing on the image feature information to obtain image identity features;

[0008] Perform re-identification on a small group of people based on image identity features to obtain the recognition result.

[0009] Furthermore, the step of constructing a semantic segmentation map from the group image, performing repair and generation processing, and obtaining the sample image specifically includes:

[0010] Obtain the group image through the camera system and input it into the image semantic segmentation model for segmentation processing to obtain the semantic segmentation map;

[0011] Perform repair processing on the semantic segmentation map to obtain the repaired image;

[0012] Perform generation processing on the semantic segmentation map based on the spatial layout generator to obtain a new spatial layout image;

[0013] Input the repaired image and the new spatial layout image into the splicing module to generate the sample image.

[0014] Furthermore, the step of performing repair processing on the semantic segmentation map to obtain the repaired image specifically includes:

[0015] Perform splitting processing on the semantic segmentation map to obtain the background mask and the pedestrian mask;

[0016] Perform point-by-point multiplication processing on the background mask and the group image to obtain the background image;

[0017] Perform repair processing on the background image and the group mask based on the depth image repair model to obtain the repaired background map;

[0018] Perform repair processing on the pedestrian mask and the group image based on the pedestrian encoder-decoder structure to obtain the repaired pedestrian cutout;

[0019] Integrate the repaired background map and the repaired pedestrian cutout to obtain the repaired image.

[0020] Furthermore, the step of performing generation processing on the semantic segmentation map based on the spatial layout generator to obtain a new spatial layout image specifically includes:

[0021] According to the semantic constraint conditions, perform position sampling processing on the semantic segmentation map to obtain the new pedestrian position information;

[0022] Obtain the corresponding coordinate information, semantic information, original position information, and angular distance information between pedestrian positions based on the new pedestrian position information, and integrate to obtain the new spatial layout image;

[0023] The angular distance information between pedestrian positions is calculated from the new pedestrian position information and the original position information.

[0024] Further, the step of inputting the repaired image and the new spatial layout image into the splicing module to generate a sample image specifically includes:

[0025] According to the heuristic rule, perform scaling processing on the new spatial layout image to obtain a scaled pedestrian cut-out image;

[0026] Perform splicing and judgment processing on the repaired image and the new spatial layout image;

[0027] When it is determined that the overlapping area of pedestrians in the image splicing part is less than the preset threshold, output the sample image.

[0028] Further, the step of constructing an information model to perform feature extraction processing on the sample image to obtain image feature information specifically includes:

[0029] Extract some sample images for vertical cutting processing to obtain pedestrian region images;

[0030] Perform expansion processing on the pedestrian region image to obtain an expanded pedestrian region image;

[0031] Perform cutting processing on the sample image through the expanded pedestrian region image to obtain sub-images;

[0032] Based on the layout feature extractor, perform feature extraction processing on the sample image and the sub-images to obtain layout feature information;

[0033] Based on the appearance information extractor, perform feature extraction processing on the sample image and the pedestrian region image to obtain appearance feature information;

[0034] Integrate the layout feature information and the appearance feature information to obtain image feature information.

[0035] Further, the step of performing fusion processing on the image feature information based on the bidirectional feature propagation module to obtain the image identity feature specifically includes:

[0036] Perform adjustment processing on the appearance feature information to obtain pedestrian appearance feature information;

[0037] Based on the layout feature information and the pedestrian appearance feature information, construct an input matrix;

[0038] Based on the bidirectional feature propagation module, calculate the input matrix to obtain the image identity feature.

[0039] Further, the calculation formula of the group feature is as follows:

[0040] f group = K·Softmax(Q1H·Q2H)·Q3H

[0041] In the above formula, fgroup Indicates the group feature, K indicates a matrix with only the first row being 1 and other rows being 0, Q1, Q2, and Q3 indicate the implicit change matrices, H indicates the input matrix, and Q1H and Q2H respectively indicate the query matrix and the key matrix after transforming the input matrix.

[0042] Furthermore, the calculation formula of the personal feature is as follows:

[0043] f person = M i ·Softmax(Q1H·Q2H)·Q3H, i = 1, 2, …, N p

[0044] In the above formula, M i indicates a matrix with only the (i + 1)-th row being 1 and other rows being 0, and N p indicates the number of people in the group image.

[0045] The second technical solution adopted by the present invention is: A small group of people re-identification system based on bidirectional feature propagation, including:

[0046] A construction module, configured to construct a semantic segmentation map according to the group image and perform repair and generation processing to obtain a sample image;

[0047] An extraction module, configured to construct an information model to perform feature extraction processing on the sample image to obtain image feature information;

[0048] A fusion module, based on the bidirectional feature propagation module, performs fusion processing on the image feature information to obtain image identity features;

[0049] An identification module, based on the image identity features, performs small group of people re-identification to obtain an identification result.

[0050] The beneficial effects of the method and system of the present invention are: The present invention performs generation processing on the group image through a spatial layout generator, so that the group image has a diverse spatial layout. By constructing a semantic segmentation map and performing repair processing on the semantic segmentation map, the sample size of the group image is expanded, the inductive bias caused by too little training data is reduced, and the generalization ability of the model is improved. By using the bidirectional feature propagation module to extract the image identity features of the group image, the recognition accuracy of small group of people re-identification for the group image is improved. Description of the Drawings

[0051] Figure 1 is the step flow chart of a small group of people re-identification method based on bidirectional feature propagation of the present invention;

[0052] Figure 2 is the structural block diagram of a small group of people re-identification system based on bidirectional feature propagation of the present invention;

[0053] Figure 3 It is a schematic flow chart of the method for processing group images based on the spatial layout generation module of the present invention. Detailed implementation manner

[0054] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. For the step numbers in the following embodiments, they are only set for the convenience of explanation and illustration, and no limitation is imposed on the order between steps. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0055] Refer to Figure 1 , the present invention provides a small group of people re-identification method based on bidirectional feature propagation, and the method includes the following steps:

[0056] S1. Construct a semantic segmentation map from the group image and perform repair and generation processing to obtain a sample image;

[0057] S11. Obtain the group image through the camera system and input it into the image semantic segmentation model for segmentation processing to obtain a semantic segmentation map;

[0058] Specifically, the crowd is captured by a pre-deployed camera system or directly input by the user to obtain a group image. The group image is input into the image semantic segmentation model to obtain a semantic information map representing the semantic information corresponding to each pixel position through pixel values. According to these pixel values, the pixel regions of each pedestrian are segmented. At the same time, the part outside the pedestrian region is used as the background region to obtain a semantic segmentation map.

[0059] S12. Perform repair processing on the semantic segmentation map to obtain a repaired image;

[0060] S121. Perform splitting processing on the semantic segmentation map to obtain a background mask and a pedestrian mask;

[0061] Specifically, refer to Figure 3 , split the semantic segmentation map to obtain a background mask and a pedestrian mask corresponding to each pedestrian. The background mask is a binary image with the same size as the original image. The value of the background mask is related to whether the position is a background region. When the value is 1, it represents that this part of the original image is a background region, and when the value is 0, it represents a non-background region. That is, the background mask has the same size as the group image, where the value corresponding to the background region is 1 and the value corresponding to other regions is 0.

[0062] S122. Perform point-by-point multiplication processing on the background mask and the group image to obtain a background image;

[0063] Specifically, multiply the background mask and the group image to obtain the corresponding background image.

[0064] S123. Repair the background image and the group mask based on the depth image repair model to obtain the repaired background image;

[0065] Specifically, the depth image repair model takes the background image and the group mask as inputs, and uses a neural network to complete the missing textures at non-background positions. The Deepfillv2 depth image repair model trained on the Places2 dataset is used. This model takes the image and the mask of the area to be repaired as inputs. The mask to be repaired is the same as the pedestrian mask. The image and the mask to be repaired are input into the model and pass through multiple convolutional layers and deconvolutional layers to obtain a logical and crowd-free repaired background image.

[0066] S124. Repair the pedestrian mask and the group image based on the pedestrian encoder-decoder structure to obtain the repaired pedestrian cutout;

[0067] Specifically, the group image and the pedestrian mask are simultaneously input into the pedestrian encoder-decoder structure trained on the Market1501 dataset. The dilated mask is obtained through a graphics dilation operation. For each pedestrian mask, after performing the graphics dilation processing on it, the dilated mask is subtracted from the pedestrian mask to obtain the mask to be repaired. The image and the mask to be repaired are input into the model and pass through multiple convolutional layers and deconvolutional layers to extract the possible holes and missing parts, obtaining the mask of the missing parts. The group image and the mask of the missing parts are spliced, and through multiple convolutional layers and deconvolutional layers, the neural network repairs the missing parts according to the surrounding information to obtain the repaired pedestrian cutout.

[0068] S125. Integrate the repaired background image and the repaired pedestrian cutout to obtain the repaired image.

[0069] S13. Generate a new spatial layout image based on the spatial layout generator for the semantic segmentation map;

[0070] S131. Perform position sampling processing on the semantic segmentation map according to the semantic constraint conditions to obtain new pedestrian position information;

[0071] Specifically, input the semantic segmentation map into the spatial layout generator composed of a neural network to generate a new spatial layout. First, the spatial layout generator generates a new position for each pedestrian. This position is randomly selected from all pixel regions in the image first, and then it is judged whether this position complies with the semantic constraints. The semantic constraint conditions are specifically that according to the semantic segmentation map, the semantic information corresponding to the sampled position cannot be an unreasonable area such as the sky or a wall without climbing tools.

[0072] S132. Obtain the corresponding coordinate information, semantic information, original position information, and angular distance information between the pedestrian positions based on the new pedestrian position information, and integrate them to obtain a new spatial layout image;

[0073] S133. The angular distance information between the pedestrian positions is calculated through the new pedestrian position information and the original position information.

[0074] Specifically, calculate the new angular distance information between the pedestrian positions. The angular distance information includes the coordinate information of the new pedestrian positions, the semantic information corresponding to the new pedestrian positions, the original position information corresponding to the new pedestrian positions, and the angular distance information between the new pedestrian positions, etc., to form a new spatial layout, generate a new spatial layout image, and the spatial layout generator can generate multiple candidate new spatial layouts.

[0075] S14. Input the repaired image and the new spatial layout image into the splicing module to generate a sample image.

[0076] Specifically, the splicing process is that the pedestrian area to be spliced in the new spatial layout image will be filled with the repaired pedestrian image, and the black area will be filled with the repaired background image.

[0077] S141. According to the heuristic rule, perform scaling processing on the new spatial layout image to obtain a scaled pedestrian cutout;

[0078] Specifically, input the repaired image, the new spatial layout image, and the depth map into the image stitching module to generate a sample image. For each candidate new spatial layout, first, according to the depth map and the new spatial layout information, follow the heuristic rule constructed based on the perspective principle of "objects closer appear larger and those farther appear smaller". The module extracts the original depth and new depth of each pedestrian according to the position information in the spatial layout information, calculates the scaling size corresponding to each pedestrian cutout, and the scaling size is proportional to the ratio of the new depth to the original depth, and perform a scaling operation on the pedestrian cutout to obtain a scaled pedestrian cutout.

[0079] S142. Perform splicing and judgment processing on the repaired image and the new spatial layout image;

[0080] S143. When it is determined that the overlapping area of pedestrians in the image splicing part is less than the preset threshold, output the sample image.

[0081] Specifically, the splicing module uses the scaled pedestrian cutout to replace the pixels at the corresponding positions in the logical crowd-free image to obtain a sample image. Judge whether there is no overlap or the overlapping area of pedestrians in the generated image. If the judged result is less than the preset threshold, it is considered that this set of positions is reasonable, store the data augmentation image and its corresponding new spatial layout information in a file, otherwise, discard this candidate new spatial layout, and finally output the sample image.

[0082] S2. Construct an information model to perform feature extraction processing on the sample images to obtain image feature information;

[0083] S21. Extract some sample images for vertical cutting processing to obtain pedestrian region images;

[0084] S22. Perform expansion processing on the pedestrian region images to obtain expanded pedestrian region images;

[0085] S23. Perform cutting processing on the sample images through the expanded pedestrian region images to obtain sub-images;

[0086] Specifically, in order to model the layout information, this step splits the input group images to generate two groups of images: the overall image and the sub-images; the overall image is directly obtained from the group images. For the sub-images, input the group images, and this step cuts the group images into M partially overlapping sub-images, thereby generating the cut sub-images. Specifically, for vertically splitting the image into M strip-shaped regions, take the regions for expansion, and cut the group images according to the expanded regions to form partially overlapping sub-images, and output the sub-images.

[0087] S24. Perform feature extraction processing on the sample images and the sub-images based on the layout feature extractor to obtain layout feature information;

[0088] S25. Perform feature extraction processing on the sample images and the pedestrian region images based on the appearance information extractor to obtain appearance feature information;

[0089] S26. Integrate the layout feature information and the appearance feature information to obtain image feature information.

[0090] Specifically, through a layout information extractor composed of, such as, a convolutional neural network, the overall image obtains a global layout representation, while the sub-images obtain a local layout representation. To solve the problem of different image sizes, flatten the obtained feature representations and pass them through a neural network, such as a Transformer structure, to obtain a consistent feature representation. Input the overall image and each pedestrian region into an appearance information extractor composed of a convolutional neural network, etc., to obtain appearance feature information, and integrate the layout feature information and the appearance feature information to obtain image feature information.

[0091] S3. Perform fusion processing on the image feature information based on the bidirectional feature propagation module to obtain image identity features;

[0092] S31. Perform adjustment processing on the appearance feature information to obtain pedestrian appearance feature information;

[0093] Specifically, a group of multi-layer perceptron networks takes the original apparent feature information as input, performs information propagation between dimensions, and uses a bidirectional feature propagation module to fuse the layout feature information and the apparent feature information of the image. The fusion result is used as the identity feature information of the group image, and the module simultaneously adjusts the apparent features to be the pedestrian appearance feature information.

[0094] S32. Based on the layout characteristic information and the pedestrian appearance feature information, construct an input matrix;

[0095] Specifically, based on the global group layout feature X GL , the local group layout feature X LL , and the member appearance feature X A , construct the input matrix H as follows:

[0096] H = [X L , X A T

[0097] X L = [X GL , X LL T

[0098] In the above formula, X GL represents the global layout feature information of the overall image, X LL represents the local layout feature information of the sub-image, X A represents the pedestrian appearance feature information, and H represents the input matrix;

[0099] Further calculate the propagation matrix, and the calculation formula of the propagation matrix is as follows:

[0100] W = Q1H · (Q2H) T

[0101] In the above formula, H T represents the transpose matrix of the input matrix, W represents the propagation matrix, and Q1 and Q2 represent the change matrices hidden in the neural network.

[0102] S33. Based on the bidirectional feature propagation module, calculate the input matrix to obtain the image identity feature.

[0103] Specifically, perform two transformations on the input matrix and multiply them, perform information propagation between the layout characteristic information and the pedestrian appearance feature information, and act on H and W through the change matrices Q1, Q2, and Q3 hidden in the neural network. According to the following formula, obtain the feature f group after bidirectional propagation and fusion. This feature is used as the group feature. At the same time, we can calculate f person , and this feature is used as the personal feature;

[0104] The calculation formula of the group feature is as follows:

[0105] f group = K·Softmax(Q1H·Q2H)·Q3H = K·Softmax(W)·Q3H

[0106] In the above formula, f group represents the group feature, K represents a matrix with only the first row being 1 and other rows being 0, Q1, Q2, and Q3 represent implicit change matrices, H represents the input matrix, and Q1H and Q2H respectively represent the query matrix and the key matrix after converting the input matrix;

[0107] The K is used to screen the group feature. Q1H and Q2H respectively represent the query matrix and the key matrix after converting the input matrix. After multiplying the query matrix and the key matrix and passing through the Softmax function, a two-way propagation matrix between H components is obtained and multiplied by the value matrix represented by Q3H. The implicit change matrix expands the expression ability of the input matrix, making the group representation more flexible;

[0108] The calculation formula of the personal feature is as follows:

[0109] f person = M i ·Softmax(Q1H·Q2H)·Q3H, i = 1, 2, …, N p

[0110] In the above formula, M i represents a matrix with only the (i + 1)-th row being 1 and other rows being 0, which is used to screen the personal feature, and N p represents the number of people in the row of the group image;

[0111] where K and M i have the same size as Q1H and Q2H.

[0112] S4. Perform small group person re-identification based on the image identity feature to obtain the recognition result.

[0113] Specifically, perform small group person re-identification on the group image with image identity feature information, and output the small group person re-identification result with higher recognition accuracy.

[0114] Refer to Figure 2 , a small group person re-identification system based on two-way feature propagation, including:

[0115] A construction module, configured to construct a semantic segmentation map and perform repair and generation processing to obtain a sample image;

[0116] An extraction module, configured to construct an information model to perform feature extraction processing on the sample image to obtain image feature information;

[0117] A fusion module that performs fusion processing on image feature information based on a bidirectional feature propagation module to obtain image identity features;

[0118] An identification module that performs re-identification of a small group of people based on the image identity features to obtain an identification result.

[0119] The content in the above method embodiments is applicable to the system embodiments of the present invention. The functions specifically implemented by the system embodiments of the present invention are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those of the above method embodiments.

[0120] The above is a specific description of the preferred embodiment of the present invention. However, the present invention is not limited to the described embodiment. Those skilled in the art can make various equivalent deformations or substitutions without departing from the spirit of the present invention, and these equivalent deformations or substitutions are all included within the scope defined by the claims of this application.

Claims

1. A small group of people re-identification method based on bidirectional feature propagation, characterized in that It includes the following steps: Construct a semantic segmentation map based on the group image and perform repair and generation processing to obtain a sample image; Construct an information model to perform feature extraction processing on the sample image to obtain image feature information; Based on the bidirectional feature propagation module, perform fusion processing on the image feature information to obtain an image identity feature; Based on the image identity feature, perform re-identification of a small group of people to obtain an identification result; The step of the constructed information model performing feature extraction processing on the sample image to obtain image feature information specifically includes: Extract part of the sample image for vertical cutting processing to obtain a pedestrian area image; Perform expansion processing on the pedestrian area image to obtain an expanded pedestrian area image; Perform cutting processing on the sample image through the expanded pedestrian area image to obtain sub-images; Based on the layout feature extractor, perform feature extraction processing on the sample image and the sub-images to obtain layout feature information; Based on the appearance information extractor, perform feature extraction processing on the sample image and the pedestrian area image to obtain appearance feature information; Integrate the layout feature information and the appearance feature information to obtain image feature information; The step of performing fusion processing on the image feature information based on the bidirectional feature propagation module to obtain an image identity feature specifically includes: Perform adjustment processing on the appearance feature information to obtain pedestrian appearance feature information; Based on the layout feature information and the pedestrian appearance feature information, construct an input matrix; Based on the bidirectional feature propagation module, perform calculation on the input matrix to obtain an image identity feature.

2. The small-group person re-identification method based on bidirectional feature propagation according to claim 1, wherein The step of constructing a semantic segmentation map based on the group image and performing repair and generation processing to obtain a sample image specifically includes: Obtain the group image through the camera system and input it into the image semantic segmentation model for segmentation processing to obtain a semantic segmentation map; Perform repair processing on the semantic segmentation map to obtain a repaired image; Based on the spatial layout generator, perform generation processing on the semantic segmentation map to obtain a new spatial layout image; Input the repaired image and the new spatial layout image into the splicing module to generate a sample image.

3. The small-scale population re-identification method based on bidirectional feature propagation according to claim 2, wherein, The step of performing repair processing on the semantic segmentation map to obtain a repaired image specifically includes: Perform splitting processing on the semantic segmentation map to obtain a background mask and a pedestrian mask; Perform point-by-point multiplication processing on the background mask and the group image to obtain a background image; Based on the depth image repair model, perform repair processing on the background image and the group mask to obtain a repaired background image; Based on the pedestrian encoder-decoder structure, perform repair processing on the pedestrian mask and the group image to obtain a repaired pedestrian cut-out image; Integrate the repaired background image and the repaired pedestrian cut-out image to obtain a repaired image.

4. The small group person re-identification method based on bidirectional feature propagation according to claim 3, wherein, The step of performing generation processing on the semantic segmentation map based on the spatial layout generator to obtain a new spatial layout image specifically includes: According to the semantic constraint conditions, perform position sampling processing on the semantic segmentation map to obtain new pedestrian position information; According to the new pedestrian position information, obtain the corresponding coordinate information, semantic information, original position information, and angular distance information between pedestrian positions, and integrate them to obtain a new spatial layout image; The angular distance information between pedestrian positions is calculated through the new pedestrian position information and the original position information.

5. The small group person re-identification method based on bidirectional feature propagation according to claim 4, wherein The step of inputting the repaired image and the new spatial layout image into the splicing module to generate a sample image specifically includes: Performing a scaling process on the new spatial layout image according to heuristic rules to obtain a scaled pedestrian cutout; Performing splicing and judgment processing on the repaired image and the new spatial layout image; When it is determined that the overlapping area of pedestrians in the image splicing part is less than a preset threshold, outputting the sample image.

6. The small-scale population re-identification method based on bidirectional feature propagation according to claim 5, wherein, The image identity features include group features and individual features, and the calculation formula of the group features is as follows: f group = K·Softmax(Q1H·Q2H)·Q3H In the above formula, f group represents the group feature, K represents the moment where only the first row is 1 and the other rows are 0, Q1, Q2, and Q3 represent the implicit change matrices, H represents the input matrix, and Q1H and Q2H represent the query matrix and the key matrix obtained after transforming the input matrix, respectively.

7. The small group person re-identification method based on bidirectional feature propagation according to claim 6, characterized in that The calculation formula of the individual features is as follows: f person = M i ·Softmax(Q1H·Q2H)·Q3H, i = 1, 2, …, N p In the above formula, M i represents a matrix where only the (i + 1)-th row is 1 and the other rows are 0, and N p represents the number of people in a row in the group image.

8. A small group of people re-identification system based on bidirectional feature propagation, characterized in that, A method for re-identifying a small group of people based on bidirectional feature propagation as described in claim 1, which includes the following modules: A construction module for constructing a semantic segmentation map based on a group image and performing repair and generation processing to obtain a sample image; An extraction module for constructing an information model to perform feature extraction processing on the sample image to obtain image feature information; A fusion module for performing fusion processing on the image feature information based on a bidirectional feature propagation module to obtain image identity features; An identification module for performing re-identification of a small group of people based on the image identity features to obtain an identification result.

Images