Abstract:A continuous obstacle avoidance strategy for a small multi-rotor UAV based on a monocular camera is proposed. The proposed method includes two modules of depth estimation and navigation decision. Among them, the conditional confrontation network is used in the depth estimation module to train and preprocess the RGB images collected by the UAV, and the deep deterministic policy gradient (DDPG) algorithm is used in the navigation decision module to realize the continuous obstacle avoidance of the UAV. On this basis, the Actor network in DDPG is improved, and the multi-modal network is used to replace the original policy network, thereby suppressing the flight vibration of the UAV and improving the obstacle avoidance ability. Finally, the test is carried out in the Airsim simulation environment. The experiment shows that the proposed algorithm model can make the UAV successfully avoid obstacles and reach the designated target point after training. Compared with the original algorithm, the obstacle avoidance trajectory has been significantly improved.