Research on pedestrian re-identification method in terminal scenarios

This paper focuses on using re-identification method to solve the problem of passenger cross-camera tracking in terminal. The terminal is usually crowded with pedestrians and the lighting keeps changing, this can cause pedestrians to be partially obscured and their color characteristics to constantly change. The existing re-identification models overly rely on the information of pedestrian appearance and edge change, and ignore the color itself, difficult to adapt to the terminal scene. In this paper, we perform feature auto-correlation analysis on the fully-connected layer of the re-identification network and extract color vectors from feature representations. By defining the dynamic changing areas of pedestrians and constructing the matching feature functions for passengers in the terminal, the problem of inaccurate matching caused by color variations and occlusions is solved. Experiments at Guangzhou Baiyun and Chengdu Shuangliu International Airports show the method outperforms DeepSort and SOLTDER, increasing Top1 Accuracy by 5.05% and Recall by 2.62%, demonstrating strong adaptability in terminal environments.