Multi-layer Joint Classification-Metric Deep Learning for Top View Image Person Re-identification

Abstract

Person re-identification (ReID) is the problem of identifying pedestrian images among multiple non-overlapping cameras. Most of the researches are based on the front view or side view of the human body. However, there are many challenges, such as occlusion, posture changes. Therefore, based on top view images, we propose a person reidentification method, called ML-CM (multi-layer joint classification-metric deep learning). The proposed method uses classification models and integrates them into the Siamese network, which is the unification of classification and metric learning. It is more suitable for top view images. In this way, minimum intra-class variance and maximum inter-class variance are achieved. We have collected the top view images of 1008 people as TVPR (top view person re-identification) dataset and conducted experiments, demonstrating that the deep features obtained by our approach can get a competitive result.