Human interaction recognition framework based on interacting body part attention

Human activity recognition in videos has been widely studied and recently gained significant advances with deep learning approaches; however, it remains a challenging task. In this paper, we propose novel framework that simultaneously considers both implicit explicit representations of human interactions by fusing information local image where the interaction actively occurred, primitive motion posture individual subject’s body parts, co-occurrence overall appearance change. change, depending on how parts each interact other. The proposed method captures subtle difference between different using interacting part attention. Semantically important other objects are given more weight during feature representation. combined attention-based representation descriptor full-body change is fed into long short-term memory to model temporal dynamics over time single framework. experimental results five used public datasets demonstrate effectiveness recognize from videos.