R-CNN and YOLOV4 based Deep Learning Model for intelligent detection of weaponries in real time video

The security of civilians and high-profile officials is of the utmost importance and is often challenging during continuous surveillance carried out by security professionals. Humans have limitations like attention span, distraction, and memory of events which are vulnerabilities of any security system. An automated model that can perform intelligent real-time weapon detection is essential to ensure that such vulnerabilities are prevented from creeping into the system. This will continuously monitor the specified area and alert the security personnel in case of security breaches like the presence of unauthorized armed people. The objective of the proposed system is to detect the presence of a weapon, identify the type of weapon, and capture the image of the attackers which will be useful for further investigation. A custom weapons dataset has been constructed, consisting of five different weapons, such as an axe, knife, pistol, rifle, and sword. Using this dataset, the proposed system is employed and compared with the faster Region Based Convolution Neural Network (R-CNN) and YOLOv4. The YOLOv4 model provided a 96.04% mAP score and frames per second (FPS) of 19 on GPU (GEFORCE MX250) with an average accuracy of 73%. The R-CNN model provided an average accuracy of 71%. The result of the proposed system shows that the YOLOv4 model achieves a higher mAP score on GPU (GEFORCE MX250) for weapon detection in surveillance video cameras.