Real-time abnormal behavior detection in indoor IoT environments using multistage YOLO, DeepSORT, and SlowFast network integration

This paper presents a real-time, top-down framework for abnormal human-behavior detection in indoor IoT environments that integrates YOLO for person localization, DeepSORT for identity-consistent tracking, and a SlowFast action recognizer. The action module uses SlowFast-R50 (4×16) pre-trained on Kinetics-400 and is fine-tuned on YOLO+DeepSORT cropped person clips, improving discrimination of safety-critical events such as medical conditions and mutual interact actions. Evaluated on a composite testbed (SPHAR, NTU RGB+D, IITB Corridor, URFD-augmented falls), the fine-tuned detector boosts average FPS from ~20 to 30.94 (a 54.7% increase) and reduces latency by 45.8% while maintaining high mAP@0.5 of 0.92; the fine-tuned SlowFast on YOLO-extracted ROIs enhances Top-1 accuracy to 84–88% across abnormal classes compared to full-frame baselines. The end-to-end pipeline sustains ≈25 FPS on a single mid-range GPU (RTX-class) and remains deployable on edge devices. Compared with pretrained or non-integrated baselines, our multi-stage approach improves abnormal-event F1 by 8–12% without sacrificing latency. By providing a reliable early-warning signal to IoT control and notification services, the system enables rapid triage and response in smart homes, clinics, schools, and other indoor facilities.