Research On Indoor Autonomous Positioning And Visual Object Tracking Algorithm Of UAV

Drones have attracted public attention for their small size,high flexibility,low cost and easy maintenance.Drones usually rely on GPS to provide positioning information outdoors,but when drones work in forests,tunnels,indoor scenarios with weak GPS o r no GPS signal,they will not be able to obtain their own location information and therefore can’t fly autonomously.In practice,many scenarios cannot ensure that the GPS signal meets the requirements of the task,e.g.logistics warehouse management,indoor sports competition live broadcast,shooting actors in the studio,as well as non-contact tasks such as patrol and publicity in the enterprise community during the epidemic.Accomplishing these tasks requires not only that drones fly autonomously to the designated locations without relying on GPS,but also that the UAV to be able to find the specific target set randomly and track it.In this thesis,we design a scheme that achieve autonomous flight and tracking of designated targets for UAV,aiming at solving the problem that UAV cannot obtain reliable positioning information in indoor environment and cannot continuously track arbitrary targets.According to the characteristics of UAV system,we try to improve the indoor positioning algorithm as well as the visual object tracking algorithm.This scheme uses an ultra-wideband positioning system to provide a reliable location information for the drone without relying on GPS.After detecting candidate targets using the visual detection,the target to be tracked can be selected through the ground station.The image collected by the monocular camera controls the UAV’s direction according to the output of the visual tracking algorithm,which continuously tracks the target.The research object of this thesis is UAV indoor positioning algorithm based on ultrawideband positioning system and single target visual tracking algorithm based on deep learning.In order to make the UWB positioning system meet the practical requirements of this project,starting from the processing of sensor ranging value,linear fitting environment error,outlier detection and elimination,Kalman filter are used to remove Gaussian white noise to achieve a more stable ranging value.Because the trilateral positioning method is not suitable for UAV positioning in motion,the extended Kalman filter algorithm is selected to fuse the ranging value to the UAV positioning value in this p roject.In order to reduce the z-axis error,the laser radar with higher accuracy is used to simplify the positioning problem of UWB positioning system from three-dimensional to twodimensional.In order to obtain more accurate initial state value,the Gau ss Newton method is used for iterative solving.In order to improve the tracking performance of the visual tracking algorithm,the existing single-target visual tracking algorithm Siam RPN is investigated.Starting from improving the robustness of the algorithm,the algorithm is improved from solely relying on the high-quality initial frame target image as the tracking template to multi frame target feature fusion,using the features extracted from initial frame and the previous frame of the target,and the historical fusion frame as the feature to fusion.It takes the fusion result of the fusion network as the tracking template of the current frame.Secondly,by introducing the residual strategy,so that the fusion result will not deviate from the initial frame information too far.At the same time,it uses a deeper network to improve the feature extraction ability of the tracker,and when the template information is fused,it uses the weighted fusion of the shallow large-resolution shape feature and the deep target semantic feature to make the extracted target feature more abundant.To further optimize the quality of the next frame tracking template information.In order to verify the feasibility of the above scheme,a four-rotor UAV system equipped with monocular camera,as well as a set of ultra-wideband positioning system are built.The positioning information output by the positioning system is used for autonomous drone and visual target tracking.