Design And Implementation Of Multi-rotor UAV Docking System On Rugged Surface Based On Deep Learning

multi-rotor unmanned aerial vehicle（UAV）has simple structure and easy operation.With the support of modern electronic technology,it has been widely used in various fields of production and life.At the same time,because the working environment is more and more complex and the remote control signal is easy to be disturbed,it is very necessary for the multi-rotor UAV to have the ability to perform tasks independently.For fully autonomous multi-rotor unmanned aerial vehicle,landing on a previously unknown area is a challenging task.When the multi-rotor UAV needs to land on an area with unknown terrain,most of the existing landing methods need manual intervention and control.In order to realize the autonomous landing of fully autonomous unmanned aerial vehicle,a fully autonomous landing system based on deep learning multi-view stereo three-dimensional reconstruction is proposed in this thesis.The main contributions of this thesis are as follows:（1）In order to make full use of the information provided by the onboard sensors of multirotor UAV,we propose a new design framework of rugged surface docking system.The framework is based on deep learning convolution neural network for multi view stereo（MVS）3D reconstruction.The rough surface docking system obtains multiple RGB images of the terrain of the landing area in the cruise of multi-rotor UAV,then extracts the Speed-Up Robust Features（surf）and executes the Structure From Motion（SFM）algorithm to obtain the sparse feature point cloud of the landing area.By using the initial depth map generated by sparse feature point interpolation and applying the fast depth 3D reconstruction algorithm Patchmatch Net-A,we can generate accurate and stable depth estimates.Subsequently,the landing area is searched based on the depth estimation value,the final docking point is selected,and the landing operation is performed.As far as we know,this is the first time to apply the learning based 3D reconstruction algorithm to the unprepared rugged terrain reconstruction on multi-rotor UAV.（2）Considering the load limitation of multi-rotor UAV and the limitation of calculation speed of onboard processor,in order to improve the calculation speed of output depth map of landing system,we propose a fast depth 3D reconstruction algorithm called Patchmatch NetA.The algorithm uses the initial depth map generated in the step of recovering the structure from motion to predict the depth value in three different resolutions,and finally provides an accurate and robust depth map.Compared with other 3D reconstruction algorithms based on depth learning,our method not only reduces the running time,but also outputs the results with little difference.The time required to output each frame of depth map is reduced from 982 ms to 736 ms.（3）In order to improve the performance of Patchmatch Net-A,we designed a new activation function called Adjustable-arctangent Linear Units（ALU）.By applying the activation function to the multi-scale feature extraction network in Patchmatch Net-A,we can improve the accuracy and robustness of feature points.After comparing with some classical activation functions,we find that ALU can improve the accuracy without changing the network structure.（4）In order to verify the effectiveness of the proposed rugged surface docking system,we used M210v2 multi-rotor UAV as a platform to carry out actual flight experiments,and successfully realized full autonomous landing.