Resnet Video 3D for Gait Retrieval: A Deep Learning Approach to Human Identification

Gait, the distinctive way a person walks, is a useful biometric trait for various applications such as crime prevention, forensic identification, and social security. Gait retrieval, which aims to find the person who matches a given gait, is an active research area, its research has drawn a significant increase. However, learning discriminative temporal features from gait data is difficult due to the subtle variations in the spatial domain of the silhouette. Recent deep learning methods have demonstrated their effectiveness for gait retrieval by learning more robust features from raw video data. In this paper, we propose a baseline network based on ResNet video R3D-18, which can capture both spatial and temporal information from the data, to address the gait retrieval problem. Our experimental results show that our optimized backbone network can extract powerful vector representations of gait and achieve high performance in retrieving the person who matches the gait from the database. On CASIA-B dataset, we obtain Rank-1 accuracy of 97.09% and Rank-10 accuracy of 99.27% under normal walking condition. The source code will be available at.