Two-Stream Convolutional Network Extracting Effective Spatiotemporal Information for Gait Recognition

Abstract

Gait recognition identifies a person based on gait feature which is a kind of unique biometric feature that can be acquired at a distance and needn’t cooperation. Gait features consist of abundant temporal features and spatial features. To make good use of the spatiotemporal information in gait features, we propose a two-stream network for gait recognition. In the temporal stream, we insert M3D architecture to an 2D network to capture the temporal information of different time perception domains. What’s more, we combine triplet loss, center loss with ID loss as our loss function to reduce the intra-class distance while increasing the inter-class distance which aids in classification. Our proposed method achieves a new state-of-the-art recognition accuracy in the CASIA-B database with the average rank-l accuracy of 95.63% on the NM subset, 90.86% on the BG subset and 72.15% on the CL subset.