A Lightweight Deep Learning-based Weapon Detection Model for Mobile Robots

Abstract

As mobile robotics continues to advance, the need for adequate surveillance in robotic environments is becoming increasingly important. Detecting suspicious objects in sensitive areas using mobile robots is challenging due to the limited computational resources available on these devices. This paper describes a new system for automatically detecting weapons in real-time video footage designed for low-computing devices in mobile robots. We present a novel weapon detection model that aims to balance the trade-off between inference time and detection accuracy, making it a lightweight model compared to existing models. The proposed model is trained and tested on existing benchmark datasets. The model is compared to existing lightweight weapon detection models to determine its suitability for low-computing devices. We obtain the mAP of 90.3%, 85.13% and 92.38% for the IITP_W, Handgun and Sohas datasets, respectively. The results outperforming the well-known PicoDet model. We envisage that the proposed model could be a useful tool for surveillance using mobile robots during events such as riots and anti-terrorist operations.