부산대학교 도서관

Based on current situation around the world, there is major need of automated visual surveillance for security to detect handgun. The objective of this paper is to visually detect the handgun in real time videos. The proposed method is using YOLO-V3 algorithm and comparing the number of false positive and false negative with Faster RCNN algorithm. To improve the result, we have created our own dataset of handguns with all possible angles and merged it with ImageNet dataset. The merged data was trained using YOLO-V3 algorithm. We have used four different videos to validate the results of YOLO-V3 compared to Faster RCNN. The detector performed very well to detect handgun in different scenes with different rotations, scales and shapes. The results showed that YOLO-V3 can be used as an alternative of Faster RCNN. It provides much faster speed, nearly identical accuracy and can be used in a real time environment.