Hybrid Silhouette-skeleton Body Representation for Gait Recognition

Gait recognition is a promising video-based long-distance biometric technology which needs no cooperation from subjects. Actually, the existing gait recognition can be roughly divided into two categories: silhouette-based and skeleton-based approaches. The former shows low computational cost but is relatively sensitive to different variations such as coat-wearing, carrying status and camera viewpoints. The latter is more robust against environmental changes, but highly dependent on accuracy of skeleton. To address the above issue, this paper proposes a hybrid silhouette-skeleton body representation for gait recognition. Inspired by the success of pose estimation approaches, we adopt recent state-of-the-art pose estimator to obtain robust body skeletons from RGB images. Besides, a novel fusion strategy is proposed to integrate silhouette and skeleton information into a single compact silhouette-skeleton representation for gait. And then we prove that heatmap for joints achieve better performance than the other skeleton representation. Extensive experiments on CASIA-B show that the proposed approach essentially boosts performance of both set-based and sequence-based gait recognition solution. In addition, it outperforms GaitPart by 1.2%, 0.6% and 2.8%, in NM, BG, and CL conditions, respectively.