Technology Of The Vision System Of Autonomous Unmanned Aerial Vehicle For Target Search

Rotor-type drones have been widely used due to their low cost,easy maintenance and strong expandability.Currently,in most application scenarios,manual operation of the drone is required.It is necessary to let the drones have the ability to perform some tasks autonomously in order to save manpower and reduce barriers to using it.In the process of autonomously performing tasks by drones,many actions can rely on visual information to be completed independently.This thesis focuses on a drone that can independently perform target search,return flight and precise landing.This thesis has done the following main work:First of all,this thesis completed the design of the vision system that assists the drone to achieve the above functions.For the tasks to be carried out by the drone,the system is divided into two parts: the ground end system and the drone end system.The hardware structure and software structure of the vision system are designed.On this basis,the workflow of the system is given.Subsequently,the learning strategy of the object detection model YOLOv3 that implements the target search function is adjusted.The model was trained on the dataset of the drone perspective.Through the experiment,the performance of object detection with YOLOv3 under the perspective of drone is analyzed.The feasibility of the scheme is verified by experimental results.The trained YOLOv3 model was optimized for deployment in the vision system.Then,in order to improve the target search function more flexible and easier to expand when changing the detectable object category,this thesis studied the method,incremental learning,of flexibly increasing the detectable object category of the YOLOv3 model.The effect of the designed YOLOv3 incremental learning method was analyzed through experiments.The experimental results show that the incremental learning method we designed can achieve a flexible and fast task expansion of YOLOv3 to a certain extent.Finally,a monocular vision positioning method based on a combination of visual fiducial markers is designed.This method enables the vision system to position the landing position with high precision over a wide range of heights.The method assists the drone to achieve precise autonomous landing during the return process.Through the experiment,the positioning accuracy of the positioning method of our visual marker combination is analyzed.Through the experiment,the positioning accuracy of the positioning method of the visual fiducial markers combination designed by us is analyzed.The experimental results show that the proposed positioning method can reduce the oscillation during the positioning process compared with the previous methods.