BioPose-based 3D Kinematic Feature Fusion for Multi-person Abnormal Behavior Recognition in Skeleton-based Video Analysis

Multi-person abnormal behavior recognition in crowded surveillance scenes remains challenging due to occlusion-induced noise in estimating 3D poses. Conventional deep learning approaches often process raw skeletal data that violates human anatomical constraints, leading to high false positive rates. To address this, this study proposes a novel framework utilizing BioPose, a biomechanically constrained optimization layer that enforces bone length constancy and valid joint limits using neural inverse kinematics. Unlike standard methods, our approach rectifies kinematic artifacts before feature extraction. We further introduce a biomechanically accurate kinematic feature fusion mechanism, which synthesizes rectified spatial coordinates with derived temporal dynamics, specifically, joint angular velocity and kinetic energy via a multi-stream adaptive graph convolutional network (MS-AGCN). Experimental results on the ShanghaiTech and CUHK Avenue benchmarks demonstrate the superiority of this approach, achieving an AUC of 87.6 and 93.5, respectively. The proposed method significantly reduces the equal error rate (EER) to 8.4, confirming that integrating physical plausibility into deep learning architectures enhances robustness against environmental noise and occlusion.