A Sequence-based Multi-Scale Network for Cross-View Gait Recognition

Abstract

Gait is recognized as a suitable feature for long distance person identification. Although horizontal partition has been proved an effective strategy for gait recognition, the existing methods do not learn the part-level features separately. In this paper, we designed a Sequence-based Multi-Scale Network (SMSN) to extract discriminative features. In addition, a multi-branch learning strategy is proposed for extracting multiple semantic at different scales. This new designed network takes the unordered gait sequence as input, and then attentive temporal pooling (ATP) method is used to measure the quality of each silhouette, and aggregates the frame-level features into sequence-level features simultaneously. The soft-max loss is further added to constrain the sequence-level features, which can improves the feature extraction ability of each branch, and reduce the difficulty of convergence. In CASIA-B dataset experiment, we achieved an average recognition rate of 96.1% under normal walking condition, which surpass the state-of-the- art methods.