Autonomous Tracking Of Quadrotors Based On Deep Learning

Quadrotor is a small unmanned aerial vehicle developed rapidly in recent years,which has been widely used in civil aerial market.The intelligence of quadrotor is one of the hot research fields at present.It mainly focuses on both autonomous tracking and obstacle avoidance at the control level.Among them,autonomous tracking refers to the aircraft in a variety of environments to automatically track the moving target meanwhile keeping a certain distance with the moving target.To achieve this purpose,most existing methods still adopt traditional robot control methods but are so limited for lack of learning ability and due to high dependence of manual adjustment.This thesis explores how to complete autonomous tracking of quadrotors via deep learning.Deep learning is an important breakthrough in the field of artificial intelligence,and can automatically extract multi-level features from training inputs.Here we exert this merit of deep learning in quadrotor to promote the intelligent level of the aircraft.To address this issue,we first design a simulator for quadrotor,named Crossline Drone(CLDrone).It consists of a controllable simulated quadrotor,a simulated car and a realistic simulation environment.Empirical studies show that it has a strong versatility and can be tailored for other intelligent aircrafts.By CLDrone,to evaluate the flight control tracking methods and train deep learning model,we build up a database including the sensing input and the pair-wise speed output by using traditional feedback control method to track the AprilTag marker.Actually,robot control output is continuous,not only related with recent sensing input but also relevant with recent short-term control output.Then,we develop a feedback neural network(FNN)model to combine two above types of historical information together,which is tailored for the robot control by taking short-term historical information as the input to improve training efficacy.Unlike the recurrent neural network(RNN),FNN takes into account the continuous control output in line with human intuition.We employ our constructed database to train FNN in a supervised way,which enables the quadrotor to track the marker based on the neural network model.Simulation results show that FNN outweighs both MLP and RNN in terms of tracking performance.To improve the tracking versatility and eliminate the dependence on the marker,we train a vehicle position estimation network based on the YOLO object detection algorithm,and then concatenate it into the trained FNN to construct an end-to-end network model from the visual perception to the control.Extensive simulation results show that the end-to-end neural network model can automatically track the vehicle by merely using visual information,and its tracking performance is superior to traditional feedback control methods.