DualGait: A Model-Based Approach to Gait Recognition using Joint and Edge-to-Vertex Dual Graph Representations

Gait recognition has gained substantial attention as a key research topic within the Computer Vision field in the past years. It consists of the re-identification of individuals based on soft biometrics such as their gait cycles referring to walking patterns. Existing approaches include appearance-based methods using silhouettes, and model-based methods relying on human joint representations. These methods face numerous challenges, such as variations in clothing and viewing angles, which affect their performance and limit their generalizability. In this paper, we introduce DualGait, a novel model-based approach for gait recognition that leverages a Spatio-Temporal Graph Convolutional Networks (STGCN) architecture enhanced with modular attention blocks. In contrast to traditional approaches, our method relies on two human representation models: a graph that captures the spatial coordinates of joints, and its corresponding edge-to-vertex dual, which encodes the lengths of body segments. We evaluate our framework on the CASIA-B benchmark, demonstrating competitive performance in gait recognition and its adaptability to variations in view angles and clothing.