A Novel Method for Ensuring Public Safety through Crowd Collision Detection Using CCTV Videos

Abstract

This paper proposes a novel approach for identifying collisions within crowds without relying on standard libraries, offering a custom methodology for video surveillance and human suspicious behavior analysis. Due to the lack of existing datasets focused on crowd collisions and suspicious movements, we collect our own dataset from diverse sources, primarily CCTV footage. In the proposed method, the dataset is processed using YOLOv8 for person detection, which efficiently segments and identifies individuals. To detect crowd collisions, the algorithm calculates the Intersection Over Union (IOU) between bounding boxes of detected people, flagging frames where overlaps exceed a threshold. The entire process, including converting videos to CSV format, filtering frames with multiple individuals, and visualizing detected collisions, designed to optimize speed and accuracy. Results demonstrate the method's success in detecting collisions in various environments, from sports fields to public spaces, showing potential for use in public safety and crowd management systems. Processing time and resource requirements are analyzed to evaluate scalability and real-world applicability