Research On Vision-based Autonomous Navigation Technology Of UAV

Unmanned aerial vehicle technology has been a hot research topic in the aviation field in recent years.With the continuous development of avionics technology,information communication technology and intelligent technology,and the increasingly mature application of deep learning technology in imagery,the field of drones has made rapid progress.At the same time,with the improvement of hardware computing performance,it is possible to carry computing resources on drones for deep learning calculations in real time.Aiming at the problem of occlusion or disappearance of the target during the real-time tracking of the drone by the drone,which leads to the problem of mission failure.Therefore,this project aims to use the drone as a platform,adopt a detection-based tracking strategy,and use airborne computing resources to monitor the camera.The captured image information is used for target detection,realizing the real-time tracking function of the unmanned aerial vehicle on the detected target,which improves the accuracy of the tracking performance compared with the traditional tracking effect.The research work of this paper is as follows: First,determine the overall technical plan according to the autonomous navigation tasks of the drone,including the modeling,selection,construction and testing of the drone.Then analyze the flight performance of the UAV to achieve the function of smooth flight.Secondly,according to the task requirements,analyze the target detection algorithm,use the powerful feature extraction capabilities of deep learning,and select the YOLOv3-tiny algorithm with simple network structure,small calculation amount,and high real-time performance.This algorithm will be accurate and fast.It is well integrated and easy to transplant.It is suitable for loading on fast-moving platforms such as drones.Then,design a controller to control the drone flight,combine traditional PID and fuzzy control,use the position information of the target detected by the camera as the input of the controller,and calculate the control signal to control the drone flight through the controller.Finally,use ROS and open source flight control PX4 for simulation verification,make the data set required for detection on the built platform,and modify the network configuration file according to the category to be detected and complete the training.Deploy the trained detection algorithm and control algorithm to the UAV platform in the simulation environment for experiments.The experimental results show that the accuracy of the detection algorithm is over 96%,and the error in the pixel plane of the tracking process is about 40,and the tracking performance is good.