Anomaly Detection in CCTV Environments through Homography-Based Speed Estimation

In this study, we propose the approach to anomaly detection for surveillance videos (CCTV) based on speed normalization and adaptive thresholding. First, the speed of each object is calculated from the sequence of bounding boxes and then converted from pixels per second (px/s) to meters per second (㎧) using homography transformation. Next, an adaptive thresholding method based on the statistical distribution of object speeds is applied to automatically determine anomaly thresholds for each camera, instead of relying on fixed thresholds. The system is implemented using YOLOv8 combined with ByteTrack, operating in real time, and provides intuitive visualization of the results directly on the video. Experiments on real-world dataset demonstrate that the proposed approach significantly reduces the rate of false alarms and effectively detects abnormal behaviors such as fast running or unauthorized vehicle entry, highlighting its potential applications in intelligent surveillance systems, autonomous driving, and public safety monitoring.