UAV Visual Odometry And Its Integration With Inertial Navigation

At present,the navigation system of unmanned aerial vehicle(UAV)relies heavily on global navigation satellite system(GNSS),which makes it unable to work normally in the GPS-denied environment,such as city,jungle or the maliciously interfered environment.Therefore,visual navigation system has become a new focused areas in navigation in recent years because of the advantages,including high positioning accuracy,high autonomy,high anti electromagnetic interference and so on.For the problems which include poor robustness of traditional visual odometry(VO)in dynamic scene and poor real-time how the algorithm operates on UAV platform,the improved semi-direct visual odometer algorithm integrated YOLOv5 is proposed.Further,the higher-accuracy morestable algorithm which integrates the aboved modified visual odometry with inertial navigation is proposed because of the problems which are both no scale initialization and no backend optimization.The main studies in this paper are summarized as follows:(1)An improved object detection algorithm which is used to integrate with visual odometry is proposed.For the problems,including the characteristics of UAV visual data,the limited computing power in edge computing platform and the strict real-time of the system,YOLOv5 object detection algorithm is used as the basic structure and integrated with the convolutional block attention module.On the premise of the realtime ensured,the algorithm improves the accuracy of small object detection in the perspective of UAV;(2)An improved semi-direct visual odometry algorithm which introduces YOLOv5 is proposed.For the limited computing power in UAV platform,the semi-direct visual odometry(SVO)is selected as the basic structure of the imodified algorithm.Bacauce of the poor robustness of dynamic scene,our proposed algorithm combines SVO and YOLOv5 for detecting the object in the scene and eliminating the dynamic feature points.Therefore,the robustness of the algorithm is improved in dynamic scene.Based on the above,the pose estimation method of initialization and the pose optimization method are modified in order to increase in the accuracy and the realtime of visual odometry.The simulation experiment used the TUM dataset is made to evaluate the robustness and effectiveness on the modified visual odometry;(3)An algorithm which combines the above visual odometry with the inertial navigation is proposed.Because of both scale-free initialization and no backend optimization on monocular visual odometry,the above improved algorithm is integrated with inertial navigation.The pre-integrated algorithm is used to process the inertial sensor data,and the loosely coupled method is used for union initialization.The tightly coupled method based on sliding window is used to optimize the position and attitude in order to further improve the accuracy of the navigation system.A quad-rotor UAV experimental platform is built,and then the vision sensor and IMU carry out union calibration.Finally,outdoor flight experiment is done for evaluating the stable and robustness of the modified visual-inertial odometry.