Physical Violence Detection in Videos Using Keyframing
J. Bineeshia, Dhanush Ram A, B. Vinoth Kumar, G Chidambaram · 2023
Smart surveillance has got tremendous traction in recent years. An important component in smart surveillance is physical violence detection. All existing violence detection models have a pre-processing step where required frames are extracted from a video. The existing interval sampling method is simple and the frames it picks sometimes lose the action in the video. In this literature we have introduced a novel key framing method based on encoding and clustering as a pre-processing step for frame extraction. In the violence detection model, we have used ResNet50 as image encode. Our model is trained on Mix dataset. A dataset formed by combining Hockey fights, Movies and Violent flows dataset. Our model is compared with start-of-the-art models on a test dataset. The test dataset is formed by collecting 20 videos from YouTube and annotating manually. Our key framed model performs better than the state-of-the-art models even with 360 frames less for a 1 minute video clip.