Detection Of Offense And Generating Alerts Using Ai

Abstract

In today’s urban environments, ensuring public safety has become a growing priority. However, traditional surveillance systems dependent on continuous human monitoring often struggle with delayed threat recognition and response. This paper presents a real-time, dual-mode surveillance system that integrates deep learning-based visual analysis with automated alert generation to enhance situational awareness in public and private security domains. The proposed system leverages a lightweight CNN (Convolutional Neural Network) trained to detect high-priority criminal offenses, namely harassment, theft, and burglary, from both live camera feeds and uploaded video fragments. A calibrated decision logic module filters out low-confidence predictions, significantly reducing false positives while maintaining high recall. To support real-world deployment, the technique integrates an alerting mechanism comprising real-time alarms, email notifications with frame evidence, and a live web dashboard for visual analytics. The lightweight design is containerized and optimized for edge deployment on devices such as the NVIDIA Jetson Nano, or mid-tier GPUs are suitable for deployment. Empirical evaluation on a composite dataset combining UCF-Crime, HarX, Shoplift-23, and proprietary CCTV clips demonstrates a classification accuracy of 92.4% and an F1-score of 89.9%, outperforming baseline models including YOLOv5 + DeepSORT. Designed with modularity, scalability, and ethical AI considerations, this research bridges the gap between theoretical computer vision models and practical, real-time crime detection solutions for smart surveillance environments.