A Cross-view Gait Recognition Method Based On Multi-scale Feature

Abstract

Gait recognition is an effective technique for long-distance person identification. However, since human gait features are affected by spatio-temporal as well as perspective, improving the cross-view gait recognition rate is still an extremely challenging task. In this paper, we propose a cross-view gait recognition method with multi-scale features(GaitMSF), which introduces temporal features of different scales through the multi-scale feature extraction module, and significant spatio-temporal clues can be captured by processing the features of different time scales. On the one hand, it introduces the relationship modeling between multi-scale features to adaptively enhance and extract important macro features and suppress unimportant features; on the other hand, it extracts short-term micro-movement features through the partial feature extraction module and achieves more effective gait recognition through the mutual complementation of short-term micro-movement features and global time features. The method is well validated on the CASIA-B dataset, achieving rank-1 accuracies of 97.7%, 93.6%, and 82.6% under normal walking (NM), carrying a bag (BG), and wearing a coat (CL) conditions.