AI-based Video Processing for Fight Detection in Colleges

The protection of civilians has always been a central issue, and, oftentimes, real-time detection of violence can help in diffusing a potentially major conflict. The Fight No Fight system is described herein as being the new calibre in fight detection and monitoring, analysing video materials through machine learning and deep learning. The system utilizes YOLO (You Only Look Once), a next-generation object detection model, on the hooting of any aggressive action shown in video streams. The system is therefore created as a web app based on Flask, allowing users to log in, upload videos, and view detected incidents in an easy-to-use dashboard. Doing so allows re-enforcement of the YOLO model’s fine-tuning to ensure high accuracy for real-time fight detection by processing video frames and sending out live alerts. All detected incidents are stored within a relational database, timestamped, geo-tagged effectively, and confidence-scored to allow easy tracking and identification. The system employs parallel processing with adaptive thresholding for efficient performance to minimize false positives while integrating an API for connectivity with legacy security architecture. Future upgrades will incorporate sophisticated action recognition models that will distinguish amongst different types of aggression and sound-activated cues such as yelling detection to refine the model. Coupled with this will be an AI Analytics dashboard, allowing the generation of reports indicating fight behavioral patterns and supports predictive security measures for proactive decision-making.