Real-Time Crime Detection System

Abstract

Traditional surveillance systems rely on human operators who can monitor only 4-6 video feeds simultaneously, with detection accuracy dropping from 95% to 70% within the first hour. This paper presents a Real-Time Crime Detection System using 3D Convolutional Neural Networks to automatically identify criminal activities in live video streams. The system processes video at 30 FPS with 91ms latency, detecting 14 crime types including fighting, robbery, assault, shoplifting, and arson. We achieved 82.3% accuracy on the UCF-Crime dataset, with the system reducing human monitoring workload by 76% and improving detection speed by 340%. Real-world deployment in a retail store over 30 days demonstrated 83% true positive rate and prevented ₹87,000 in losses.