Enhancing Surveillance System through Video Anomaly Detection

The rapid proliferation of surveillance infrastructure in urban environments necessitates intelligent systems capable of autonomous monitoring and real-time threat assessment. Traditional surveillance frameworks relying on manual operator oversight are susceptible to cognitive fatigue, scalability constraints, and delayed response times. This paper presents an AI-based intelligent surveillance system integrating three functional modules: deep learning-based person detection using YOLOv8, multi-object tracking via DeepSORT, and cross-camera person re-identification (ReID) employing convolutional neural network-based feature extraction. The proposed end-to-end pipeline maintains identity consistency across non-overlapping camera views under challenging conditions including occlusion, illumination variability, and significant pose changes. Experimental evaluation on a custom multi-camera dataset demonstrates a mean Average Precision (mAP@0.5) of 91.4% for detection, a Multi-Object Tracking Accuracy (MOTA) of 82.5%, and a Rank-1 re-identification accuracy of 87.2% on the Market-1501 benchmark. The complete pipeline sustains an average throughput of 22.3 FPS on mid-range GPU hardware, confirming real-time operational feasibility. The system establishes a robust, extensible foundation for next-generation intelligent surveillance with planned extensions toward behavioral anomaly detection and automated alert generation.