DyGait: Gait Recognition Network Based on Skeleton Dynamic Features

Abstract

In the field of gait recognition, research mainly revolves around binary silhouette and skeleton structures based on joint points. While existing gait recognition network models have shown promising results in simple indoor environments, they still encounter substantial challenges in complex outdoor settings. Currently, most methods employ a simplistic approach to extract dynamic information, typically by subtracting adjacent frames. However, we believe this method oversimplifies the process and fails to effectively capture essential dynamic features.To address this limitation, we propose a novel method called DyGait,a gait recognition network based on skeleton dynamic features. Before conducting feature extraction, we introduce a dynamic feature streaming technique to establish associations between long-term dynamic information, as illustrated in Figure 1. Furthermore, we enhance feature extraction by incorporating motion captors to better capture global motion changes.In our experiments, we comprehensively compare DyGait with various current gait recognition methods across multiple public datasets. The results demonstrate that DyGait consistently outperforms other methods, irrespective of indoor or outdoor scenarios.

Keywords

• Gait recognition
• dynamic information extraction
• graph convolution
• indoor/outdoor scenarios
• skeleton structures