Multi-Level Fine-Tuned Transformer for Gait Recognition

Gait Recognition aims to identify individuals by their walking patterns. Abundant spatial and temporal characteristics are contained in human walking for indicating unique personal identity. There have been a large number of existing gait recognition approaches achieving good performance via the analysis of spatial and temporal characteristics. However, most of them directly concatenate the separately extracted temporal features and spatial ones for gait recognition, rarely considering the overall spatiotemporal structure and the inner connection between them. Therefore, we establish a Multi-Level Fine-tuned Transformer trained on sequential gait data, to capture and analyze the overall spatiotemporal information for gait recognition. Specifically, we first design a point-level spatial feature extractor for spatial feature extraction of frames of the gait sequence, based on typical Transformer architecture with positional encoding and multi-head self-attention mechanism. Subsequently, a frame-level temporal feature extractor is employed for the overall spatiotemporal features by learning dynamic information of the extracted spatial features from frames. Moreover, fine-tuned strategy is followed in the process of model training, to profoundly explore intra-frame spatial information, as well as global spatiotemporal features of the whole walking sequence. Extensive experiments conducted on three public datasets (normal and abnormal gait) show promising classification results, compared with several state-of-the-art techniques.