Abnormal Behavior Detection using Object Detection and Tracking
Mark Denzel Becina, Dionis A. Padilla · 2022
Video surveillance and computer vision have been integrating rapidly. A full-pledged detection surveillance system, requiring proprietary hardware and systems, is uncommon in a regular household or small business. This study aims to bring person tracking, running detection, and fall detection into a highly modifiable package to make computer vision more accessible. The system utilizes the person detection capabilities of YOLOv4 in combination with the DeepSORT tracking algorithm to track people in a frame and derive useful metrics that would aid in detecting running and falling behavior. The accuracy tests on the system showed 78% on running detection and 80% on fall detection. FPS testing shows that the system could run at an average of 7.4 frames per second on entry-level hardware. The system utilizes a web-based approach with an API and an included web UI to show the customizability of the project.