Research And Application Of Autonomous Localization And State Estimation Methods For UAV

The UAV’s autonomous navigation technology in GPS-denied environment is a key technology in the field of UAV.It enables UAVs to complete complex tasks such as autonomous navigation,mapping,planning and autonomous obstacle avoidance in an unknown complex environment without GPS.The basis of the autonomous flight of UAV is to accurately obtain the flight status through the onboard sensor.In recent times,there has been an explosion of research on visual Simultaneous Localization and Mapping(SLAM)due to the unique advantages of vision sensors.However,the visual SLAM relies on the illumination and texture of the scene and the frequency of the pose state is limited by the image frame rate,which can not satisfy the design requirements of the controller.In this thesis,we focus on the SLAM scheme of point and line features and state estimation algorithm to improve the robustness of SLAM in different scenes and provide high-frequency and reliable state information for the controller.The main research contents are as follows:Firstly,considering the limitation of the onboard computing power and the cumulative error of the traditional filtered SLAM,we analysis the stereo SLAM method based on the Multi-State Constrained Kalman Filter(MSCKF).The method uses the observations of the point features in the multi-frame images to constrain the pose of the corresponding observation frame,ensuring that the features jointly optimize the multiframe pose and reduce the cumulative error.And it also designs a front-end frame of point feature matching and tracking and a back-end filtering model using point feature re-projection error for pose state update.In addition,we propose a dynamic adjustment strategy for stereo cameras’ exposure time,which improves the robustness of SLAM to scene and illumination changes.Secondly,considering the problem that the localization accuracy of SLAM based on point feature is decreased in weak texture structured environment,we propose a SLAM method based on MSCKF using point and line features extracted from the stereo cameras.We introduce line features based on Line Segment Detector(LSD)and Line Band Descriptor(LBD)of the structured environment to the SLAM and design a framework based on point and line features for feature matching and tracking to ensure stable tracking of the point and line features.At the same time,we design a back-end filtering model which uses the point and line features’ re-projection error to constrain the observation frame pose using the orthonormal representation of the line features.It improved the localization accuracy when the point features are insufficient.Finally,considering the problem that the frequency of the SLAM pose does not satisfy the design requirements of controller and the safety of the UAV is affected when the SLAM position is highly offset,we propose a state estimation method based on Unscented Kalman Filter(UKF).We design a terrain estimator using radar to ensure the accuracy of the height estimate under various terrain conditions.In addition,we align the time of multiple sensors.We predict the state using Inertial Measurement Unit(IMU)and update the state by visual inertia SLAM and radar.In this way,we increase the output frequency of the pose state.Furthermore,we design a multi-sensor redundancy strategy for height estimation to ensure that the UAV can safely land through the height estimation of the radar when the SLAM height is estimated to be offset.