Is Multi-Person Gait Recognition Feasible under Mutual Occlusion? A Human Model Regression-based Approach

Current gait recognition studies primarily focus on identifying a single person, assuming only one subject appears in the scene. However, in the wild, people often walk in groups, leading to mutual occlusion that degrades recognition performance. In this paper, we propose Model-MultiGait, a human model regression-based approach for multi-person gait recognition that explicitly accounts for mutual occlusion. Specifically, given a gait image sequence with multiple walking subjects, we directly regress parametric human models for all individuals, and extract gait features from each model to recognize them in an end-to-end manner. Unlike the standard pipeline that first detects each subject and then fits models independently, where mutual occlusion may cause missed detections and prevent model fitting for occluded individuals, our method simultaneously regresses multiple human models, capturing spatial and temporal relationships between subjects from a holistic perspective. Experiments on synthesized multi-person gait datasets demonstrate the effectiveness of the proposed method, achieving superior performance compared to state-of-the-art gait recognition methods.