Real-Time Crime Monitoring System using Deep Learning Techniques

The increase in crimes occurring in urban areas requires a change from the manual methods of surveillance which are inefficient to intelligent automated systems. We propose in this paper a robust real-time crime monitoring system based on a multi-modal deep learning framework for live video feed analysis. Our system comprises three specialized modules: a Buffalo_L model with ArcFace for criminal face recognition, an R3D-18 architecture augmented with Channel Attention and LSTM for violence detection, and a YOLOv8 model for weapon detection. The results of these modules are combined to carry out a dynamic risk assessment, which in turn produces alert levels (Low, Medium, High Risk) and records the data. Tested against standard datasets, the system achieves high performance. Specifically, the modules for face recognition, violence detection, and weapon detection attain accuracies of 98.29%, 96.68%, and a mAP@50 of 86.11% respectively. This integrated system is a scalable, intelligent, and proactive way to increase public safety and facilitate the work of the police.