Gait recognition and re-identification based on regional LSTM for 2-second walks

Abstract

Law enforcement and different authorities need a new efficient way to track and re-identify a person of interest via different cameras. Usually, the person of interest is not known and the original video may be short and have poor quality. In this paper, we propose a new technique based on a new regional-LSTM learning model that can use a 2-second walk to recognize and re-identify an unknown person. The proposed technique first targets the rhythm of movements in different regions of the body by creating a separate LSTM model for each region. Then, outputs from 22 regions are combined in a subnetwork to extract the relations and different degrees of uniqueness of all regions. The proposed regional LSTM model creates a gait-embedded vector to represent a 2-second walk. Experimenting on imbalanced and balanced datasets, the results show that the proposed regional LSTM model performs significantly better than the existing techniques on the Cumulative Matching Characteristic (CMC) curves and top- k accuracy, Receiver Operating Characteristic (ROC) curves, and Precision-Recall (PR) curves. This indicates that the proposed technique has a high-ranking performance (CMC test), can efficiently distinguish the gaits of a subject from others (ROC test), and occupies high relevancy (PR test). From the experimental results, it is likely that one in four videos retrieved from the proposed techniques shows the person of interest with over 90.8% and 85.7% accuracies in imbalanced and balanced data, respectively. This demonstrates that the proposed regional LSTM model is efficient and useful in tracking and re-identifying a person of interest.