Gaitstc: A Pure Spatial-Temporal Convolution Network For Skeleton-Based Gait Recognition

Gait recognition is a promising biometric identification technique that utilizes video analysis to identify individuals from a distance based on their walking patterns. Conventional methods focus on analyzing gait characteristics via sequences of silhouette images. However, these silhouette representations often contain redundant information, leading to extracted features that might encompass extraneous visual cues. From a theoretical perspective, gait recognition methods based on skeletal representations discard this superfluous data, concentrating solely on the kinetic information of the human skeleton. This skeletal-centric approach inherently provides greater robustness, producing features that more accurately represent individual gait attributes. Building upon this premise, we introduce the Gait Spatial Temporal Convolution Network(GaitSTC). This method employs convolutional techniques to extract fine-grained spatial-temporal gait features, capitalizing on the restricted skeletal data by capturing salient features from each frame. Innovatively, our study incorporates convolutional methods, commonly used in appearance-based approaches, into skeletal-based techniques, resulting in a richer extraction of fine-grained gait features and highlighting the potential advantages of model-based gait recognition methods. Empirical evaluations on the widely accepted CASIA-B dataset reveal that our method outperforms previous model-based approaches and remains competitive with representative appearance-based gait recognition techniques.