Non-local Temporal Modeling for Practical Skeleton-Based Gait Recognition

Gait, a unique biometric identifier for recognizing individual identity at a distance, plays an important role in practical applications. Existing gait recognition methods utilize either a gait set or a sequence. However, these methods ignore the periodic characteristic of gait, where actions at one moment are related to actions at another moment. As a result, their recognition accuracy in real scenes can significantly decrease due to noise and frame loss. To deal with this issue, we design a NLGait network to explore the temporal relation among gait frames, which adaptively leverages both local and non-local relations to achieve practical gait recognition. Specifically, we design multi-scale temporal information extractor (MTIE) to capture these relations. Furthermore, we design an attention based adaptive frame fuser (AFF) to aggregate the features of frames in a gait sequence. Extensive experiments have verified the competitive accuracy and robustness of our method. The accuracy of the counterpart methods is degraded by 8.9% and 19.3%, respectively, due to noise and temporal loss, while ours is degraded by only 3.6% and 2.7%.