Intelligent Security Surveillance System Based on Multi-Modal Object Detection and Edge Computing

Abstract

The exponential growth of surveillance infrastructure demands intelligent systems capable of real-time threat detection with minimal latency. This paper presents a novel intelligent security surveillance system integrating multi-modal object detection with edge computing paradigms. Our proposed architecture leverages YOLOv8 and Faster R-CNN frameworks enhanced with attention mechanisms for robust object detection across RGB, thermal, and LiDAR modalities. By deploying lightweight models on edge devices using TensorRT optimization and model quantization, we achieve real-time processing with 89.7% mean Average Precision (mAP) while reducing inference latency to 47ms. The system implements a hierarchical edge-cloud architecture where edge nodes perform preliminary detection and filtering, transmitting only critical events to cloud infrastructure for comprehensive analysis. Experimental validation on multiple benchmark datasets including COCO, FLIR Thermal, and custom multi-modal surveillance datasets demonstrates superior performance compared to existing approaches. Our system achieves 94.3% detection accuracy for person detection, 91.8% for vehicle detection, and 88.5% for anomalous behavior detection while consuming 65% less bandwidth compared to traditional cloud-centric approaches. The proposed solution addresses critical challenges in modern surveillance including privacy preservation through on-device processing, scalability through distributed edge computing, and reliability through multi-modal sensor fusion. Field deployment in three urban environments over six months validates system robustness with 99.2% uptime and <50ms end-to-end latency. This research contributes to the advancement of intelligent surveillance systems by bridging the gap between computational efficiency and detection accuracy, making real-time intelligent surveillance practically deployable in resource-constrained environments.