Metric-Based Key Frame Extraction for Gait Recognition

Gait recognition is one of the most promising biometric technologies that can identify individuals at a long distance. From observation, we find there are differences in length gait cycle and quality each frame sequence. In this paper, propose novel framework to analyze human gait. On hand, designed Multi-scale Temporal Aggregation (MTA) module models temporal aggregate contextual information with different scales, on other introduce Metric-based Frame Attention Mechanism (MFAM) re-weight by importance score, which calculates using distance between frame-level features sequence-level features. We evaluate our model two popular public datasets, CASIA-B OU-MVLP. For normal walking, rank-1 accuracies datasets 97.6% 90.1%, respectively. complex scenarios, proposed method achieves 94.8% 84.9% under bag-carrying coat-wearing walking conditions. The results show top level among state-of-the-art methods.