Occlusion-Aware Human Mesh Model-Based Gait Recognition

Earlier approaches to occluded gait recognition can broadly be divided into reconstruction-based and reconstruction-free methods. Reconstruction-based techniques attempt to recover an unobstructed silhouette or gait representation before performing feature extraction and matching. Reconstruction-free techniques instead learn features directly from incomplete observations or compare only regions that remain visible in both samples. Although these strategies can be useful, many depend on a full-body bounding box for cropping, scale normalization, and body-center registration. This assumption becomes unrealistic when the complete person cannot be observed.

A model-based strategy offers a more natural alternative. By fitting a human model to the visible evidence, the system can estimate body pose, body shape, and the location of the complete body even when some regions are hidden. The proposed work extends this idea from individual images to continuous gait sequences. Temporal information is important because the visible region can change from frame to frame, and a sequence provides complementary evidence about the person's movement.