Deep Learning-Based Framework for Real-Time Object Detection in Smart Surveillance Systems

Abstract

Smart surveillance systems require rapid and reliable interpretation of continuous video streams, but conventional motion-based methods frequently produce false alarms and provide limited semantic understanding. This paper proposes a deep learning-based framework for real-time object detection in smart surveillance environments. The methodology combines video-frame acquisition, adaptive preprocessing, a one-stage convolutional detector, confidence filtering, non-maximum suppression, multi-object tracking, and event-driven alert generation. The detector is designed around the speed advantages of the YOLO family while incorporating multiscale feature extraction and lightweight deployment principles established in MobileNet and EfficientDet. A surveillance-oriented evaluation protocol is defined using precision, recall, F1-score, mean average precision (mAP), frames per second (FPS), and end-to-end latency. The proposed framework targets common surveillance classes such as persons, vehicles, and restricted-zone objects under illumination changes, partial occlusion, and crowd density. An illustrative prototype evaluation indicates that a YOLO-based configuration can provide a more favorable real-time speed-accuracy balance than a two-stage Faster R-CNN baseline, while lightweight backbones further reduce computational cost. The framework is intended as a reproducible methodology for intelligent CCTV analytics and edge-assisted surveillance applications.