IMPROVE DETECTION PERFORMANCE OF WEAPONS CARRYING BEHAVIOR

Abstract

Weapon carrying behavior detection is one of the necessary problems. This system deploys on a large space with many cameras. It is installed in a compact and slow configuration while it requires high accuracy. The YOLO network is one of the most effective object detection models in image/video. However, the published YOLO (You Only Look Once) models are not trained on weapons databases. This study collects and labels the database of 6 common types of weapons. In addition, the paper also implements a transfer learning solution between high-configuration and heavy YOLO models to lower-configuration and smaller YOLO models. This work aims to transfer the knowledge from the expert system that has compiled about weapons to a simpler model to achieve a model with a simple configuration, easy to deploy in practice while maintaining higher accuracy. The research also proposes a solution to detect the user's weapon carrying behavior that combines the output of the YOLO model with the optical flow stream of detected humans and weapons in the image. The results show that our system achieves better results when transferring learning to detect weapons and especially achieves better results when detecting weapon carrying behavior. Experimental weapon detection accuracy between YOLO V3 tiny models compared to YOLO V3 tiny – YOLO V8 KD is 2% to 7.9% higher while response time is 10.95 ms faster when tested on GPU. The weapon carrying behavior detection accuracy of the proposed solution is improved by up to 13.88% compared to using the model output of YOLO V3 tiny.