Research Of Tracking And Localization In Visual-inertial Odometry

Visual-Inertial Odometry(VIO)is a technique to recover the trajectory and attitude of an object as well as the distance marched with the help of inertial measurement units and cameras.With the rapid development of mobile robot technology,research and application of VIO become very popular.Due to the theoretical significance and practical value of VIO,we studied the tracking and positioning algorithms.The main work and research results are as follows.Firstly,to solve the problem that the inefficiencies of traditional feature point matching cannot make sure the real-time performance of VIO,an accelerated algorithm for feature matching for SIFT-like feature descriptors,and their derived descriptor is proposed.Theoretical analysis results show that the algorithm can speed up the original matching by more than seven times.In the real test,the matching speed of the algorithm to the feature points increased by two to four times.Numerous experimental results show the effectiveness and stability of the algorithm.Secondly,to improve the real-time performance of the VIO and solve the plane failure problem in the existing schemes,the paper presents the mathematical proof of the plane failure problem,analyzes the advantages and disadvantages of the iterative five-point algorithm and the classical direct five-point algorithm which is based on the linear equation,respectively.The paper also improves an iteration-base algorithm for the camera pose recovery.The algorithm recreates the two-view geometry by re-selecting the coordinate system,reducing the computation required during model solution.The empirical data shows that the improved algorithm's efficiency and anti-noise ability are better than the classical linear equation solving method.Finally,a method to fuse the inertial data and the visual data based on loose coupling is presented,and a prototype of VIO system is built for demonstration.Experimental results verify the correctness of the method and verify that the accuracy of the loosely coupled data fusion scheme is limited to a low-level.