An Improved Method Based on Deep Reinforcement Learning for Target Searching

Unmanned Aerial Vehicle (UAV), due to their high mobility and the ability to cover areas of different heights and locations at relatively low cost, are increasingly used for disaster monitoring and detecting. However, developing and testing UAVs in real world is an expensive task, especially in the domain of search and rescue, most of the previous systems are developed on the basis of greedy or potential-based heuristics without neural network. On the basis of the recent development of deep neural network architecture and deep reinforcement learning (DRL), in this research we improved the probability of success rate of searching target in an unstructured environment by combining image processing algorithms and reinforcement learning methods (RL). This paper aims at the deficiency of target tracking in unstructured environment, trying to propose an algorithm of stationary target positioning of UAV based on computer vision system. Firstly, a new input source is formed by acquiring depth information image of current environment and combining segmentation image. Secondly, the DQN algorithm is used to regulate the reinforcement learning model, and the specific flight response can be independently selected by the UAV through training. This paper utilizes open-source Microsoft UAV simulator AirSim as training and test environment based with Keras a machine learning framework. The main approach investigated in this research is modifying the network of Deep Q-Network, which designs the moving target tracking experiment of UAV in simulation scene. The experimental results demonstrate that this method has better tracking effect.