AI-Based Video surveillance System

Abstract: Rapid urbanisation and escalating public safety demands have created an urgent need for intelligent, automated surveillance solutions that can operate without continuous human oversight. Conventional CCTV infrastructure places excessive cognitive load on operators monitoring multiple feeds simultaneously, increasing the risk of missed incidents due to fatigue and delayed response. This paper proposes a dual-model AI surveillance framework that concurrently detects three real-world emergency categories — road accidents, fire incidents, and suspicious human activity — by combining YOLOv8 spatial object detection with ResNet-50 temporal activity classification in a unified processing pipeline. On emergency detection, the system autonomously assembles an alert payload containing a timestamped snapshot, GPS-tagged camera location, and event confidence score, dispatching notifications in parallel via SMS, email, and mobile push notification to relevant authorities. Experimental evaluation on publicly available benchmark datasets yields detection accuracies of 89.2%, 91.5%, and 86.0% for accidents, fire, and suspicious activity respectively, with per- frame inference latency of 0.8–1.2 seconds and end-to-end alert delivery within three seconds. The proposed framework significantly reduces reliance on manual monitoring and offers a scalable, deployable foundation for smart city infrastructure, transportation hubs, and public safety control rooms.

Keywords: Artificial intelligence; video surveillance; YOLO; CNN; emergency detection; computer vision; real-time alerts; smart cities; deep learning.