SocioGesture: Real-Time and Adaptive Social Gesture Perception for Human-Robot Interaction

The `SocioGesture` system is designed to enhance human-computer interaction, featuring real-time and adaptive social gesture recognition capabilities. The system utilizes a compact representation of the body - hand skeleton and a lightweight dual-stream model, combining limb movements with hand postures to achieve low-latency recognition on edge devices. To improve deployment robustness, the model incorporates training data containing scenarios such as missing hands, occluded arms, and time-instability key points. In tests using social gesture datasets in both indoor and outdoor scenarios, SocioGesture demonstrated strong unseen object recognition capabilities, significantly improving performance under structured joint occlusion, and successfully ran on robot-mounted edge devices. Additionally, during system deployment, uncertain interaction segments are saved for offline annotation and adaptation, thereby maintaining the original category performance while expanding the gesture vocabulary.