Autonomous Obstacle Avoidance Algorithm of UAVs for Automatic Terrain Following Application

Autonomous obstacle avoidance flight is a key capability for unmanned aerial vehicles (UAVs) in the automatic terrain following application, which ensures that the UAVs can perform complex, versatile and difficult movements in the flight. Existing obstacle avoidance methods such as the visual SLAM, generally require artificially specifying the feature values that need to be extracted and are susceptible to illumination and obstacle positions. While artificial intelligence technology has made breakthroughs in many fields, the advantages of the neural networks overcome these shortcomings. Therefore, we propose an autonomous obstacle avoidance algorithm using deep reinforcement learning method. Concretely, a virtual three-dimensional visual simulation environment is established firstly, which simulates the flight states of the UAV and output the state image in real time according to the control decision. A deep convolutional neural network is then built as the brain of the intelligent agent, which takes the state image of the UAV as input, and outputs the discretized control decision to control the UAV. Moreover, The Deep Q Network method is employed to train the convolutional neural network in an autonomous way, that is the intelligent agent can try to control the UAV to avoidance the obstacle by itself. After training, the UAV is controlled by the trained convolutional neural network to complete the autonomous obstacle avoidance task during the flight. This algorithm is realized through continuous self-learning and self-evolution, which enables the UAVs to utilize visual information in complex scenes, adjust flight height in real time according to terrain height and obstacle height. The research results will have a strong application prospects in both military and civilian areas.