Research On Vision Based Adaptive Landing Of UAV On A Moving Plattform

In recent years,due to the rapid development of machine learning and deep learning,and the continuous improvement of computer hardware,artificial intelligence has begun to appear in people’s vision.In the world’s intelligent trend,combined with the application in the field of UAV,when the UAV returns from its autonomous mission,landing on a moving platform with a large movement is an important part of it.In This thesis,we focus on this direction,combines with the current research situation of the UAV industry,and carried out the research topic of research on vision based adaptive landing of UAV on a moving platform after actual research.It mainly includes the following tasks:(1)A monocular vision target recognition algorithm and depth estimation strategy based on deep learning are designed to solve the problem that traditional drone tracking relies on artificial cooperative identification,especially at medium and long distances,the camera recognition target is limited by the size of artificial identification.The algorithm uses YOLOv3-Tiny for small target recognition and applies a 3D information calibration detection strategy to determine its relative 3D position.(2)For the long-time tracking problem,the real-time performance of traditional algorithms is poor,and the deep convolutional neural network algorithm has the characteristics of low fault tolerance.In this thesis,we propose a real-time tracking algorithm under irregular motion.This algorithm improves the TLD algorithm,by integrating the YOLOv3-Tiny algorithm in the detection module and its related processes,so that the window generation part of its detection module runs on the GPU using the detection algorithm of deep learning,another online learning module and tracking module running on the CPU,real-time performance of TLD algorithm improved through GPU and CPU hybrid processing strategy.(3)A UAV moving landing strategy is designed to work when the long-range tracking algorithm of the UAV runs near the moving target.The relative position estimation of the UAV and the target position is performed through the Apriltags recognition algorithm.Then,through the estimation of the position of the moving target and the UAV height estimation method based on complementary fusion filtering,the UAV flight trajectory is output,and a trajectory tracking controller based on model-free control of the cascade DJI black box PID is designed.(4)Construct simulation and actual systems to verify the algorithm and landing strategy proposed in this thesis.Firstly,simulation experiments were carried out through Gazebo and Microsoft’s autonomous driving simulation platform AirSim,which verified the feasibility of the visual recognition algorithm part of this thesis.Then in the real environment,the DJI Matrices 210V2 platform was used to verify the trajectory controller,and finally the UAV landed on the stationary and moving UGV.