Towards robust gait recognition integrated with human mesh model

Gait recognition is a popular biometric that can identify people based on their walking patterns, even from a distance without the subject's cooperation. Currently, most gait recognition approaches rely on human silhouette-based representations, which are easy to extract even from low-resolution images, but are sensitive to variations caused by challenging factors that often occur in real application scenarios (e.g., viewing angle, partial occlusion). For example, in the case of partial occlusion of a human body due to obstacles or limited camera field of view, existing methods require the bounding box or the height of the entire human body as a prerequisite, which is not observable in occluded scenes. In this presentation, we introduce an occlusion-aware model-based gait recognition method that works directly on gait videos under occlusion without the above-mentioned prerequisite. Specifically, given a gait sequence that only contains non-occluded body parts in the images, we directly fit a skinned multi-person linear (SMPL)-based human mesh model to the input images without any pre-normalization or registration of the human body. Experiments on occlusion samples simulated from the OU-MVLP dataset demonstrated the effectiveness of the proposed method. We then applied this method to images obtained from a fisheye camera mounted on a navigation robot, where the images contain severe image distortion and partial body occlusions. By combining with a set of preprocessing procedures, our method demonstrated outstanding effectiveness on challenging real-world scenarios. Finally, we show some extensions of constructing a gait database using human meshes, and applications to soft biometrics such as gait-based age estimation.