Air-to-air Micro Air Vehicle interceptor with an embedded mechanism and deep learning

Abstract

Illegal flight of small and portable aerial vehicles, or the so-called Micro Air Vehicles (MAVs), has become a serious safety issue in most cities around the world. Most existing intercepting systems hardly detect them due to their size, or these systems are too expensive to install everywhere. This paper presents a novel solution to tackle this problem using another aerial vehicle that is capable of flying in close proximity to intruding MAVs, detecting, and intercepting them. This vehicle is embedded with an intercepting mechanism and a vision-based detector for autonomously tracking and intercepting the target. The tilted arm design installed with inverted motors and nets maximizes the chance of intercepting or even capturing intruding MAVs. A deep learning-based detector integrated with a Kalman filter tracker is proposed to enable real-time MAV detection and tracking using an onboard processing unit and a single camera. In this study, 9332 images of flying MAVs were used for training the pre-trained model. YOLOv4-tiny showed the best balance of accuracy and inference speed among the models chosen for the analysis, which obtained an average precision of 90.04% and 14 FPS in the evaluation. In addition, a control strategy that adapts the detected MAV size or its relative distance to the MAV interceptor in the velocity controller is developed to approach the intruding MAV steadily. The performance of the proposed system for aerial target tracking was validated in several flight tests. The results show that our vehicle can successfully detect, track, and intercept the moving target.