Performance evaluation of 3D Human Pose Estimation algorithms for skeleton-based gait recognition

Gait recognition is a biometric technique that identifies individuals based on unique walking patterns. Two main categories of approaches dominate this field: silhouette-based methods, which analyze a person's body shape to extract gait features, and skeleton-based methods, which focus on modeling the person's pose to represent gait movements. Silhouette-based methods are currently preferred in applications for their higher accuracy; however, skeleton-based methods are rapidly catching up, due to advancements in Human Pose Estimation (HPE) algorithms, and represent a promising upcoming alternative to silhouette-based approaches. This paper evaluates the effect of different 3D HPE algorithms on skeleton-based gait recognition, by building a gait recognition system that combines the extraction of 3D poses from RGB video data with a contrastive attention-based gait encoder for recognition. We benchmark the performance of different HPE algorithms in our system using the CASIA-B dataset, focusing on how improvements in accuracy of 3D pose data affect the Rank-1 recognition accuracy of a gait recognition algorithm. Our findings provide an analysis of suitable 3D HPE models for this task and a new benchmark result for 3D skeleton-based methods, representing a step forward in the viability of approaches based on 3D skeletons for gait recognition.