MEN-VVDF: Multipath excitation network-based video violence detection framework focusing on human activity in keyframes

To date, video violence detection remains a challenge in visual communication because violent events are sudden and unpredictable, making it difficult to efficiently define and locate the occurrence of violence from video data. In addition, the complexity and redundancy of video limits the existing methods the ability to extract relevant information and the accuracy of detection. Thus, effectively recognizing violence from video clips is still an open problem. This paper proposes a video-level framework for constructing human action sequences and detecting violence. Firstly, a keyframe extraction algorithm is developed to capture representative and informative frames. Then, a strategy is introduced to emphasize human actions and eliminate background bias. Lastly, a novel neural network is designed to excite spatio-temporal, channel, and motion features to effectively model violence. The proposed framework is comprehensively evaluated on two large-scale benchmark datasets. The experimental results demonstrate that the proposed framework outperforms the existing state-of-the-art schemes and achieves classification accuracies of more than 98% and 94% for the two datasets.