A new drone chasing drone approach based on deep reinforcement learning with accelerated rewards

Abstract

In this paper, we propose a deep reinforcement learning-based approach that uses the drone camera as the only input source for a drone to track another drone in real-time autonomously. Deep learning and deep reinforcement learning algorithms are developed for this proposal. First, one of the object detection algorithms, YOLO, was trained with a dataset of different drone images to instantaneously detect drones in the images taken from the camera of the following drone. The detected drone was enclosed in a box and positioned on the screen. The size information of the box was sent to the agent trained with the Deep deterministic policy gradient (DDPG) deep reinforcement learning algorithm to determine the action of the following drone. This approach is tested in five different following scenarios. These two scenarios were also tested in environments with different light levels. According to the results of the scenarios, the following accuracy rate is calculated to be at most 99 %, and the drone response accuracy rate is calculated to be at most 95 %. In addition, the OpenAI Gym simulator is designed according to our problem, the DDPG agent is trained, the performance of the developed system is tested in a real-world environment, and the results are observed.