Research And Implementation Of UAV Obstacles Avoidance Method Based On Monocular Depth Estimation

Due to the advantages of small size,affordable price and stable control,multi-rotor drones have been gradually widely used in various fields in the past decade.It has also been widely studied.The obstacle avoidance of UAV Based on monocular vision is one of the challenging problems in many research issues.This research is of great significance for UAV miniaturization,intelligence,multi-sensor fusion and disaster recovery.At the same time,the self-supervised depth estimation proposed greatly reduces the difficulty of constructing the data set and training the depth estimation model in the corresponding scene,which makes the UAV obstacle avoidance based on monocular depth estimation highly possible.This paper researches the UAV obstacle avoidance based on self-supervised monocular depth estimation and implements corresponding prototype system.The main work includes the following:(1)For the concern of UAV outdoor obstacle avoidance,the forest and street scenes dataset are constructed.Based on the principle of self-supervised monocular depth estimation,combined with relevant optimization studies to build and train a relative depth estimation model.After that,the relative and absolute depth conversion relationship is established through measurement and the estimated effect is evaluated.The failure scenarios of depth estimation in drone flight samples are sorted and summarized,and the whole picture uncertainty estimation method is proposed based on the reconstruction loss to identify the estimation failure.(2)According to the hardware constraints and the depth estimation perception characteristics,the forward monocular UAV obstacle avoidance mode and the corresponding two-dimensional random environment are constructed.In this mode,available rule avoidance algorithms are designed based on human obstacle avoidance experience.Considering the shortcomings of artificial rule algorithm,a UAV obstacle avoidance policy is constructed and trained based on the PPO deep reinforcement learning algorithm and the random environment.In addition,for the situation that may fall into the local optimal area after the trained policy,a policy based temporary goal is added to increase the possibility of escape.(3)For the simulation environment,based on Airsim,build a simulation verification environment that matches the constraints of the real environment.Experiments in the simulation environment show that the obstacle avoidance method proposed in this paper can be effectively applied to UAV horizontal obstacle avoidance and navigation in 3D scenes,and the proposed reinforcement learning policy with high obstacle avoidance capability.For the real environment,according to the system requirements,a scalable perception,computing and control UAV link is established in the form of restful services,and the corresponding mobile and server applications are designed and implemented.Experiments in a real environment show that this method and the corresponding prototype system can effectively identify and avoid different types of small obstacles and move forward in the forest.