Comparison You Only Look Once (Yolo) Algorithm On Physical Violence Video Detection

Abstract

Physical violence is one of the crimes that often occurs in various environments and can have a serious impact on victims, both physically and mentally. One of the obstacles in handling it is the delay in detecting acts of violence. The solution to this problem is to implement the best algorithm between You Only Look Once (YOLO) version 8 and version 9 to detect physical violence through video automatically and quickly. The dataset used consists of two classes, namely violence and non-violence, which have gone through the process of extraction, data cleaning, and labeling using Roboflow. The model was trained using Google Collaboratory, and the training results were evaluated using mAP, precision, recall, and F1-score metrics. Based on the test results, YOLOv9 obtained the best performance with a precision of 0.8096, recall of 0.8665, F1-score of 0.8363, and mAP of 0.8117. The detection system is then implemented into a web-based application using the Flask framework, which allows users to Upload videos and detect acts of violence automatically. The test results show that the application runs according to its function and is able to detect physical violence well. This research is expected to be a supporting solution in video-based security surveillance systems.