A Deep Learning Framework for Violence Detection in Videos Using Transfer Learning

Violence detection in videos has become a significant problem in the field of computer vision. It involves the process of automatically identifying violent behavior in video content. The rapid growth of digital media led the researchers to focus on developing effective methods for detecting violence that can automatically identify real-world instances of violence in order to maintain public safety and security. This paper presents a transfer learning approach for detection of violence in videos. The approach uses a pre-trained ResNet50, a deep residual network to extract the features from frames of videos. The results show that the suggested approach achieved accuracy of 98.89% on Hockey Fight and 99.97% on Movies datasets and highlighting the effectiveness of transfer learning in learning discriminatory features for recognition of violent action in videos over traditional hand-crafted feature detectors.