Deep Learning Oriented Vision Guidance Systems For Unmanned Aerial Vehicle Path Planning And Landing

Unmanned aerial vehicles(UAVs)have been widely used in civilian,scientific and military fields.With the rapid development of computer vision technology,using the computer vision system as the UAV guidance system has become one of the most popular topics of current researches.Path planning and autonomous landing are the key technologies for UAVs to achieve autonomous flight.Current vision-based UAV path planning systems heavily rely on accurate modeling of the surrounding environment or a large number of manually labeled datasets,which requires a lot of programming work or a lot of labor costs.The vision-based autonomous landing system relies heavily on specific landing targets,while the universal detection algorithm has low accuracy and it is difficult to be adapted to diverse environments.Meanwhile,the conventional autonomous landing method requires the flat areas to land.In view of the above problems,this thesis has designed visual-based UAV path planning and autonomous landing systems based on the deep convolutional neural network.The specific research work is as follows:First,considering the problem that traditional UAV path planning methods need to model the enviroment first,this thesis adopts the deep sarsa to the UAV path planning.This thesis uses the real-time images captured by the UAV and the position relative to the final target as state information.The deep convolutional neural network is utilized to calculate the state action value function of the action taken by the UAV.The action is selected and updated by the theory of sarsa.Finally,the UAV can reach the final target position with the shortest path while avoiding obstacles.Then,considering the problem that universality algorithms do not have good detection effect on landing targets,this thesis uses the deep convolutional neural network to learn the high-level features of landing targets to improve the accuracy and robustness of the landing targets detection.In order to deploy it on the UAV's on-board microprocessor and realize realtime landing area detection,this thesis uses the methods that reduce the size of the convolution kernel,reduce the feature map dimension to the convolution layer and delay the downsampling operation.At the same time,this thesis takes the IoU error into consideration.These methods can reduce the amount of computation while ensuring the landing targets detection accuracy.Aiming at the problem that UAV needs a flat area to land,this thesis designs a bionic landing gear with soft joints.The landing gear replaces the joints of conventional claws with flexible materials to make them more like the claws of birds.The soft joint of the landing gear can effectively reduce the force during the UAV landing.What's more,as the landing gear has the ability to grasp like a claw,it can assist the UAV to land on the tree branch-like surface area.Finally,the experiments are carried out to verify the proposed vision-based UAV guidance systems presented in this thesis.The effectiveness of the path planning system is verified by AirSim simulation environment,which is very close to the real world.The effectiveness of protecting UAV during it landing is verified by ANSYS simulation.What is more,the performance of UAV landing on tree branch-like surfaces is verified by a real quadrotor with the bionic landing gear.