Research On UAV Indoor Obstacle Avoidance Based On Vision

The UAV has been widely used in crowd control,epidemic prevention publicity,body temperature measurement,express delivery and disinfection.However,due to the lack of the ability of UAV to autonomously avoid obstacles in unknown environments,the application of UAV in indoor environments with complex obstacles is limited.The traditional autonomous obstacle avoidance method based on SLAM and SFM requires the UAV to hover repeatedly,inferring the space and obstacles that the UAV can traverse by extracting and matching environmental feature points and then using conventional path planning strategies to control flight.The above methods need to be equipped with expensive environment sensing sensors such as lidar and depth camera,and has high requirements on the onboard computer.However,when faced with untextured walls and moving obstacles,the obstacle avoidance effect is not good.In this thesis,based on the characteristics of many indoor types,complex texture information,and moving and prohibited obstacles,a visual obstacle avoidance method is proposed.It consists of a deep estimation module based on unsupervised learning and a obstacle avoidance decision module based on deep reinforcement learning,details as follows:First,it explained the basic concepts of deep learning,then introduced three most commonly used network models and their characteristics,including autoencoders,convolutional neural networks,recurrent neural networks and analyzes the basic process of reinforcement learning to solve problems.Then it extended to deep reinforcement learning and introduced three classic deep reinforcement learning methods such as DQN,DDPG and A3 C.Which laid the foundation for the following UAV environment perception and obstacle avoidance decision.After that,the principles of commonly used environment-aware sensors were analyzed.In view of the shortcomings of the sparse and complex calculation of the depth map obtained by lidar and depth camera,a depth estimation method based on monocular color image was proposed.Using stereo vision principle and bilinear interpolation method,the problem of depth estimation of environmental information was transformed into the reconstruction process of input image.The use of unsupervised learning methods reduced the reliance on real depth information datasets during model training.The image reconstruction loss,depth smoothing loss and depth consistency loss functions were set asmodel training loss functions as needed.The experimental results show that compared with the existing depth estimation method,the depth estimation method proposed in this thesis is more convergent,the model training time is shorter and the accuracy of the depth estimation in this thesis is higher under the four evaluation indicators such as root mean square error and square root relative error.Finally,studying the principle of UAV obstacle avoidance and transforming the UAV obstacle avoidance problem into a partially observable markov decision process.The depth information obtained by the color image through the above depth estimation model is used as the UAV pair observations of the environment.The design implements a deep reinforcement learning UAV obstacle avoidance control strategy.The core is to use the deep recurrent Q network with time attention to make the optimal obstacle avoidance action according to the distance between the UAV and the obstacle.The experimental results show that our proposed method enable UAV fly for a longer time before the obstacles collide compared with the existing method and can have a good autonomous obstacle avoidance flight in a simulated indoor environment with stationary and moving obstacles effect.