Autonomous Drone Control for Visual Search Based on Deep Reinforcement Learning

Autonomous flight of drone using Deep Reinforcement Learning is an attractive area of research in recent years that gives excellent results. Autonomous drone flight is defined through a set of complex tasks for understanding the environment and navigating independently through it. Understanding the environment means that the drone knows its location in respect to other objects and that it can easily reach the desired location without collision. Extending the problem with a target search task increases the complexity and the necessity for using new tools and algorithms. In this paper, we present an approach in which a drone, in addition to learning to navigate in an unknown environment, learns how to find and approach an object a priori assigned to it as a target. In our approach, the drone uses RGB and RGB-D cameras as the only source of information about environment. Our proposed solution incorporates, into the framework of deep reinforcement learning, appropriate fast object detection, feature extraction, as well as efficient existing algorithms for avoiding obstacles. The proposed model uses the sensed RGB-D image of the drone as the main factor for estimating the distance to the obstacles, while, on the other hand, our model also requires two RGB images for a Siamese network as feature extractor used to identify the target in the environment, group of these images represents the current general state, based on which drone performs the action for which it can potentially receive the highest reward. We used a 3D simulator (MS AirSim) to validate the performance of our approach. Based on the simulation results, we conclude that the proposed method exhibits promising performance in terms of the rate of successful approach to the required target.