Multi-Scale Temporal Adaptive Convolution for Gait Recognition

Gait recognition is one of the most promising biometric technologies which widely used in video security. The spatio-temporal features contained in gait sequences provide rich discriminative information for individual recognition. However, most existing gait recognition methods pay more attention to spatial features and ignore temporal information. To address this issue, we design a multi-scale temporal adaptive convolution module (MTAC) in the proposed framework. Specifically, we introduce the adaptive temporal modeling strategy to generate convolution kernel based on specific videos, and combine the adaptive temporal kernel with multi-scale dilated convolutions to extract more robust gait temporal information. We conduct experiments on the CASIA-B dataset. The proposed method achieved recognition accuracies of 97.2%, 93.5% and 83.9% under three conditions, namely normal walking and walking with a backpack or coat, respectively, showing the effectiveness and advancement of the propose method.