Development of an anti-drone system using a deep reinforcement learning algorithm

Abstract

This paper represents the study of tracking and capturing an invading drone using an vision-based drone to defend it. To achieve the goal, it is required that the tracker UAV can detect a drone first and estimate the location of the target and track it. Existing image processing is not appropriate because the size of the target drone is too small. In this research, in order to overcome these shortcomings, a deep learning-based detection algorithm was applied, which has excellent performance and robustness, and the estimation for the position of a drone using the detected information was developed. Most studies for tracking targets have used techniques a lot, usually used in the missile guidance law. However, there are many parameters that need to be tuned according to the situation and the speed of the tracking drone, so tracking performance is not constant. To complement this problem, this research proposed an algorithm that intelligently tracks by applying the deep reinforcement learning method. The framework in deep reinforcement learning can find the optimal behavior through repetitive episodes, but it takes a lot of time and cost if it is conducted as an actual experiment. In this paper, through GAZEBO simulation that reflects the drone model well, many repetitive episodes were performed to learn, and overall algorithm was verified in the simulation. Finally, a drone platform was developed, and it was verified through a tracking experiment with an actual target.