Real-Time Video Violence Detection Using CNN

Abstract

Abstract: In order to effectively enforce the law and keep cities secure, monitoring technologies that detect violent events are becoming increasingly important. In computer vision, the practice of action recognition has gained popularity. In the field of computer vision, action recognition has gained popularity. The action recognition group, however, has mainly concentrated on straightforward activities like clapping, walking, jogging, etc. Comparatively little study has been done on identifying specific occurrences that have immediate practical applications, like fighting or violent behaviors in general. The responsiveness, precision, and flexibility of violent event detectors are indicators of their effectiveness across a range of video sources. This capacity might be helpful in specific video surveillance situations. Several research focused on violence identification with an emphasis on speed, accuracy, or both while ignoring the generalizability of various video source types. In this paper, a deeplearning-based real-time violence detector has been proposed. CNN serves as an extractor of spatial features in the suggested model. Here, a convolutional neural network (CNN) architecture called MobileNet V2 is utilized to extract frame-level information from a video, and LSTM, which focuses on the three factors (overall generality, accuracy, and quick reaction) as a temporal relation learning approach.