Suspicious activity detection – using computer vision and deep learning
Venkata Rao Yanamadni, Ankam Sai Gowtam, Sujoy Kodali, Methuku Shashank Reddy · 2025
The increasing security needs and public safety demands make video surveillance systems essential to protect areas including public spaces and commercial as well as residential locations. The massive amount of video footage recorded by surveillance systems requires more resources than security personnel can manage through manual examination of all frames to spot suspicious behaviours resulting in overlooked threats. By training its CNN model on variety of labelled video set, it learns to recognize and categorize several incorrect or suspicious activities which afford it the power to notice actions like battling, theft, hearth incidents, shootings and accidents. It trains the model with 80% of data and test the model with 20%, achieving the astonishing prediction accuracy of 99.91%. The model is evaluated using important performance metrics like precision, recall, and F1 score which also help in minimizing false positive, and false negative errors along with its high accuracy. The web-based application developed using Django framework allows users to upload their videos, while incorporating the real-time activity detection functionality offered by the system. The system examines video content for specific actions and displays alerts with confidence levels that allow security operators to respond in real time.