Appearance and Gait-Based Progressive Person Re-Identification for Surveillance Systems

Person re-identification has attracted increasing research interest due to its great potential ability to find the target person in large-scale surveillance videos. Most existing methods for person re-identification only achieve limited performance in practical applications, as they mainly focus on the generic appearance of person while neglecting some unique identities of person (e.g., human gait). In this paper, we propose a progressive person re-identification approach to simultaneously improve the timeliness and accuracy of identifying the target person. Our approach is treated as a coarse-to-fine search in the feature space, which consists of appearance-based coarse filtering and human gait-based fine search. It first employs the multi-level appearance attributes of person for shrinking the search area, then exploits human gait feature to accurately identify person. To alleviate the impact of viewpoint changes on gait cycle feature extraction, a Long Short-Term Memory-based Siamese network is designed to learn view-invariant and discriminative periodic motion cues of gait sequences. Evaluations on two benchmark datasets: PRID-2011 and iLIDS-VID. Experimental results demonstrate that our method can outperform the state-of-the-art methods in both MAP and CMC curves.