Video pedestrian re-identification method based on channel attention mechanism and application

Abstract

The invention discloses a video pedestrian re-identification method based on a channel attention mechanism and application. The method comprises the steps: preprocessing a video sampling image, inputting the preprocessed video sampling image into a convolutional neural network, respectively extracting N image feature maps, and obtaining a temporary video-level feature map through time pooling; inputting the temporary video-level feature map into a channel attention module, and outputting feature vectors; expanding and filling the feature vectors in height and width to obtain an attention matrix, and multiplying the attention matrix by the original N feature maps element by element to obtain modified N feature maps; performing time and space pooling in sequence to obtain a video-level feature vector, adding a classifier to obtain a classification result, calculating cross entropy loss of the classification result, and training and optimizing the whole model; and performing pedestrian re-identification application on the trained pedestrian re-identification model. According to the method, image-level features are better fused into video-level features, the expressive force of the video-level features is increased, and the performance of video pedestrian re-identification is improved.

Description

technical field [0001] The invention relates to the technical field of video pedestrian re-identification, in particular to a video pedestrian re-identification method and application based on a channel attention mechanism. Background technique [0002] In the networking application of public security video surveillance construction, a large number of surveillance cameras are installed in public places, such as schools, hospitals, railway stations, subway stations, airports, etc. The full coverage of public security video surveillance can effectively protect the personal and property safety of the people. However, how to effectively use these large amounts of public safety surveillance video is another difficult problem. The traditional method of relying on the naked eye to search for video to match the target person is often time-consuming and has a low accuracy rate. The executives of the traditional method generally feel exhausted. Faced with the explosive growth of surv...

AI Technical Summary

Problems solved by technology

[0005] Defects or deficiencies of existing technologies: 1. The use of prior knowledge to assist in solving the problem of pedestrian re-identification is often limited by the data set, for example, in some data sets, the change of optical flow is not obvious or the parts of pedestrians in the image are not complete

In addition, the acquisition of pedestrian parts based on pedestrian images also depends on the extraction tools trained on other datasets, so the extraction tools will also restrict the acquisition of pedestrian parts

2. The spatial attention mechanism often uses the correlation between the corresponding image feature pixels as the fusion basis, and its granularity can reach the spatial size of the image feature map, such as 16*8, 24*12, and the granularity of the temporal attention mechanism It is often larger, only the number of frames N in a sampled video (generally N<=16), and the larger granularity leads to the weak ability of the attention mechanism to modify features; 3. The existing channel attention mechanism is still relatively Monotonous, such as "Co-segmentation Inspired Attention Networks for Video-based Person Re-identification" directly averages the channel attention coefficients corresponding to each image feature, and then assigns each original image feature as a Hadamard product

4. According to the image quality and then fuse the image features, the quality coefficient is directly calculated from each image feature. On the one hand, the granularity of fusion lies in the number of frames, which is larger than the spatial attention. On the other hand, only through the image itself. The attention coefficient obtained by comparing image features in non-video is difficult to use as a basis for fusion

Images