REAL-TIME CRIME DETECTION IN VIDEO STREAMS USING DEEP LEARNING AND WEB TECHNOLOGIES

Abstract

The rising rate of criminal incidents in urban regions has intensified the need for robust surveillance 
systems that can identify suspicious activities in real time. Conventional crime detection approaches 
depend heavily on manual monitoring of CCTV footage, which is not only labor-intensive but also 
susceptible to human oversight and inefficiencies, especially in large-scale environments. These 
traditional systems are limited in their capacity to autonomously recognize and react to abnormal 
events such as theft, assault, or fire, resulting in delayed responses and compromised public safety. To 
address these shortcomings, this project introduces a web-based crime detection system built with the 
Django framework, which utilizes a hybrid Deep Convolutional Neural Network (DCNN) and 
Recurrent Neural Network (RNN) model. The proposed system analyzes video input by processing 
video frames using OpenCV and classifying them into specific crime categories. With an impressive 
performance—accuracy of 99.90%, precision of 99.96%, recall of 99.97%, and F1-score of 99.91%, 
along with a minimal mean squared error (MSE) of 0.09%—the application provides highly accurate 
real-time detection. Users can upload videos via a user-friendly interface, while administrators are 
provided with a dashboard to train models and monitor system performance. This integration of deep 
learning with web technologies results in an intelligent, scalable, and responsive crime detection 
platform. Ultimately, this system reduces human intervention, enhances surveillance efficiency, and 
strengthens safety measures in public areas, offering valuable support to law enforcement and urban 
security operations.