An Improved Weapons Detection and Classification System

Abstract

The rapid increase in insecurity orchestrated by the illegal possession and usage of weapons (knives, rifles, handguns, amongst others) has led to a wider deployment of surveillance cameras for real-time video monitoring. However, the small size of these weapons, distance from the surveillance camera and atmospheric conditions have made it impossible for easier identification of the weapon or the crime perpetrators. Interestingly, several research have deployed computer vision through closed-circuit television cameras monitoring and tracking of weapons. Nevertheless, the need to improve detection accuracy, and lowering false alarm rates has remained a bottleneck. This paper addressed some of the inherent issues using a selective tile processing strategy that uses an attention mechanism. The image tiling technique was adopted as weapon images are smaller in size when compared to the entire image and down-sampling the images to a lower resolution will either reduce the features of the small weapon or make it invisible to be detected. Hence, the high-definition images from the surveillance camera in the public Mock Attack dataset were automatically splited into their respective tile images using their respective size ratio to the input of the modified capsule network. The capsule network was adopted for detecting and classification of the system owing to speed in prediction, lower data requirement, ease in pose recognition, texture, and image deformation. An average accuracy, precision, recall, and F1-score of 99.43%, 98.14%, 98.77%, and 98.45% respectively was achieved.