Generalised robotic anomaly detection in dynamic public environments

Abstract

This paper presents a real-time anomaly detection system integrated into a ROS-enabled mobile robot for public safety monitoring in dynamic environments such as shopping centers. The system targets three critical anomalies: fallen individuals, abandoned bags, and visible knives. Our final approach combines YOLO-World v2 for object detection, YOLO-Pose for posture estimation, and GPT-4V for contextual reasoning. In controlled and public scenarios, the system achieved 88.3% accuracy and high precision (fall: 95.6%, knife: 91.7%), improving on the state of the art in fall detection recall (79.6% vs. ~49.7%). We detail the architecture, deployment strategy, and performance evaluation, and discuss latency, lighting sensitivity, and privacy challenges.