Research On Positioning Technology Of UAV In Indoor Dynamic Environment Based On Visual-inertial Fusion

With the application fields of unmanned aerial vehicles(UAV)being further extended,the autonomous navigation and positioning of UAV has become increasingly important especially in the indoor environment where the UAV could face challenges of satellite signal loss and disturbance of dynamic features.This paper focuses on a tightly coupled Visual Inertial Odometry(VIO)which operates for UAV’s positioning in the indoor dynamic environment.On the premise of satisfying the real-time quality of the system,our work aims at identifying dynamic features to reduce their interference and using the visual information to reduce the cumulative error from the IMU.The main contents are as follows:Firstly,the calibration algorithm is studied to accurately integrate the measurement information from the camera and IMU.Considering the defects of traditional camera calibration methods such as the loss of dimensional information and unstable calibration results,an improved 3D calibration algorithm is proposed to achieve more accurate and stable camera internal parameters.Meanwhile,the methods of calibrating IMU internal parameters and camera-IMU external parameters are studied to provide accurate calibration parameters.Secondly,the tightly coupled VIO is mainly focused.In the visual front end,a dynamic feature recognition method based on motion segmentation is proposed,which uses IMU measurements to perform motion compensation and calculates pixel differences.The dynamic features in the environment can be identified and removed by using the improved K-means algorithm to cluster features according to the different pixel differences between the static and dynamic features.The measurements from the camera and IMU are tightly coupled in the framework of Extended Kalman Filter(EKF).In order to reduce the complexity of calculation,camera poses are used as the states of the system instead of features and the states are updated by establishing multi-state geometric constraints.The method of camera state amplification during the filtering is improved to reduce the impact of the cumulative error from IMU,by using the fusion result of the vision and IMU as the new camera pose in the state vector.Thirdly,an experimental UAV platform is built mainly based on the NVIDIA TX2 core development board on which the experiment is carried out.By collecting data in the actual dynamic indoor environment,the performance of the dynamic feature recognition algorithm can be evaluated and the trajectory of the flight can be estimated.Compared with the original MSCKF algorithm,the superior accuracy of the VIO proposed here is verified.The simulation on datasets and real flight experiment results show that the VIO algorithm proposed in this paper can accurately distinguish dynamic features in the environment and can obtain accurate pose estimations which are comparable to the VINS＿Fusion in real time.Compared with the classic MSCKF method,the VIO proposed here can better reduce cumulative errors and improve the accuracy of the whole system.