Combining Keyframes and Image Classification for Violent Behavior Recognition

Surveillance cameras are increasingly prevalent in public places, and security services urgently need to monitor violence real time. However, the current violent-behavior-recognition models focus on spatiotemporal feature extraction, which has high hardware resource requirements can be affected by numerous interference factors, such as background information camera movement. Our experiments have found that violent non-violent video frames classified deep-learning models. Therefore, this paper proposes a keyframe-based scheme. scheme considers independent events judges based whether number of keyframes exceeds given threshold, reduces requirements. Moreover, overcome we propose new training method background-removed original image pair facilitates extraction does not add any complexity networks. Comprehensive demonstrate our achieves state-of-the-art performance for RLVS, Violent Flow, Hockey Fights datasets, outperforming existing methods.