Establishment method and application of foreground-guided and texture-focused pedestrian re-identification model

Abstract

The invention discloses a method for establishing a pedestrian re-identification model based on foreground guidance and texture focusing and its application, and belongs to the field of pedestrian re-identification. The local branch containing the BottleNeck module is introduced after the last feature layer of the basic network, and the attention branch containing the foreground attention module and the texture-focused decoder are introduced after the penultimate feature layer to obtain the network to be trained; the texture-focused decoder is used As the decoder part of the network to be trained, the global branch, the local branch and the attention branch are used as the non-decoder part of the network to be trained, and the non-decoder part and the decoder part are alternately trained; after the training is completed, remove the to-be-trained The texture-focused decoder in the network is added to the feature output layer to obtain a pedestrian re-identification model based on foreground guidance and texture focus. The invention can improve the robustness and accuracy of pedestrian re-identification.

Description

technical field [0001] The invention belongs to the field of pedestrian re-identification, and more particularly, relates to a method for establishing a pedestrian re-identification model with foreground guidance and texture focusing and its application. Background technique [0002] The purpose of person re-identification is to retrieve pedestrians belonging to the same identity in camera scenes without overlapping areas, and it is a fundamental task in video image processing and computer vision. In recent years, with the development of deep learning, person re-identification has attracted more and more attention from the community and researchers, and has made considerable progress. It is used in the field of intelligent video surveillance (such as multi-target tracking, pedestrian search, etc.) Has significant application value. [0003] Pedestrian re-identification is a difficult task with many challenges, including large differences in pedestrian resolution, pose chang...

AI Technical Summary

Problems solved by technology

Pedestrians with the same identity are often in different backgrounds, while pedestrians with different identities are often in similar backgrounds. This situation is very easy to mislead the learning direction of the network, and it is difficult for the network to extract the discriminative features of pedestrians.

[0004] Some existing attention-based methods use the attention mechanism to obtain the potential semantic or structural relevance of pedestrians to highlight valuable pedestrian appearance information, but most of the attention methods face the problem of high computational overhead, and it Treating the foreground and background equally is likely to cause feature false alarms; there are also some methods that directly use image segmentation algorithms or pose estimation algorithms to determine body parts or key point positions to guide feature extractors to generate targeted pedestrian body parts. Eigenvectors, which are effective in scenes with large occlusions and pedestrian pose changes

However, the introduction of image segmentation / pose estimation as an independent module into the person re-identification task will complicate the construction of the overall model, cannot be trained end-to-end, and seriously affects the actual inference speed, and more importantly, is subject to background interference. The impact of image segmentation or key point positioning may be inaccurate, and the accuracy of pedestrian re-identification will be compromised at this time

In general, the robustness and accuracy of the existing pedestrian re-identification methods still need to be improved due to the problem of feature false alarms.

Images