Learning Local-Global Feature Representation for Pedestrian Detection and Re-Identification
Tingyi Cai, Dawei Zhang, Yiting Wang, Zheng Zheng · 2023
In recent years, exploration of pedestrian detection and Re-identification technology with the important goal of building intelligent monitoring systems has made good breakthroughs with the development of deep learning. However, in the constantly changing and complex real-world scenarios, this task still faces many challenges such as image resolution, size, perspective, pedestrian posture, etc. The algorithms are far from practical application. In this paper, we propose a deep learning-based pedestrian detection and re-identification algorithm that focuses on improving feature extraction and similarity measurement. We attach importance to the extraction and utilization of local features and train both local and global branches together. At the same time, through repeated comparative experiments, we improve the design of the metric loss function, so that the model performance compared to the initial model on Market-1501: rank-1 reaches 0.91 and mAP reaches 0.78, a relative improvement of 10.22%. In addition, our approach demonstrates competitive performance on the challenging MSMT17 dataset.