Deep learning-based human suspicious activity detection system for defence surveillance

Abstract

Threats are metamorphosing simultaneously at an alarming speed, demanding instantaneous responsiveness in defence-related surveillance applications. Conventionally, these systems involve monitoring by using rule-based and manual detection, which implies delays in detection and incorrectness at the highest rate of false positives. To mitigate such inconveniences, the present framework is intelligent enough to recognize suspicious human activities such as trespassing, fighting, or weapon handling from live video streams. The issue of rule-based systems causing delayed and incorrect detection in defence surveillance is discussed. A deep learning-based solution is introduced that utilizes the R (2+1) D 3D CNN for temporal recognition and YOLOv8 for real-time object detection, which occurs with exceedingly high precision. The DCSASS and UCF-Crime databases serve as the source of the training datasets which guarantees precision for activity detection in real-time. The 3D model outperforms 2D CNN: 98% accuracy, lower loss. Hence, the incorporation retains movement subtleties that are critical for classifying these actions while keeping an edge on the computation time. Performance validation is conducted on benchmark datasets such as DCSASS and UCF-Crime, which speak volumes about high precision, recall, and prompt response times. A comparison shows how considerably advanced this system is than the existing system. Furthermore, it is scalable for urgent deployment in sensitive operational zones.