UAV First View Landmark Localization via Deep Reinforcement Learning

In recent years, the study of Unmanned Aerial Vehicle (UAV) autonomous landing has been a hot research topic. Aiming at UAV’s landmark localization, the computer vision algorithms have excellent performance. In the computer vision research field, the deep learning methods are widely employed in object detection and localization. However, these methods rely heavily on the size and quality of the training datasets. In this paper, we propose to exploit the Landmark-Localization Network (LLNet) to solve the UAV landmark localization problem in terms of a deep reinforcement learning strategy with small-sized training datasets. The LLNet learns how to transform the bounding box into the correct position through a sequence of actions. To train a robust landmark localization model, we combine the policy gradient method in deep reinforcement learning algorithm and the supervised learning algorithm together in the training stage. The experimental results show that the LLNet is able to locate the landmark precisely.