Research On SLAM Visual Odometry Based On Feature Point Matching

With the rapid development of artificial intelligence technology,more and more intelligent devices are integrated into people's lives and industrial manufacturing.However,the importance of autonomous positioning and navigation is self-evident in order to truly realize the intelligent office of mobile devices.SLAM(simultaneous localization and mapping)technology plays a key role in solving the above problems.This technology introduces the method based on estimation theory into the robot's mapping and positioning links,so that the mobile carrier can realize synchronous positioning and map building.The vision odometry(SLAM front-end)is mainly responsible for the positioning work.Its task in RGB-D SLAM system is to analyze and process the continuous vision information obtained by the sensors mounted on the mobile device,and then provide a high-quality initial value for the back-end module,and then calculate its position and attitude and the relative movement of new road signs with the change of time.Compared with traditional odometry and monocular and stereo vision odometry,3D vision odometry based on depth camera can directly capture color map and corresponding depth map due to its inherent properties,avoiding the trouble of obtaining image depth value through related complex calculation,and greatly improving the operation efficiency of visual odometry.In this paper,the front-end module of RGB-D SLAM is studied systematically,and some improvements are made to solve the problems such as high error matching rate of feature matching algorithm and poor registration accuracy of image data facing scale change.The specific research content and improvement work are as follows:First of all,it introduces the concept characteristics of visual slam technology and the performance characteristics of each sub module.The structure model of the second generation sensor of Microsoft,the causes of image distortion and its elimination strategy are discussed.Finally,the calibration experiment is carried out based on the calibration principle of Zhang Zhengyou.The MATLAB calibration tool package is used to calculate the internal parameter matrix and estimate the external parameter value,which verifies the necessity of the calibration process.Secondly,an improved ORB algorithm is proposed to deal with the shortcomings of ORB operator in scale-oriented mutation image processing and high mismatch rate.The specific improvements are as follows:firstly,A-KAZE is used for reference to construct scale space based on nonlinear diffusion filter;secondly,FAST-9 detector of ORB is used to detect corner points on the created scale space layer;next,FLANN method is used to calculate the Hamming distance for rough matching of feature points;on this basis,two paths are used to eliminate the mismatching point pairs,One is to use the more mature PROSAC algorithm to remove noise points,The other is to establish the Poisson distribution mathematical model to carry out the secondary screening of the point pairs after the initial matching.and apply the improved results to the follow-up links;in the process of camera pose optimization,PnP fusion ICP optimization algorithm is used to ensure the effective operation of motion estimation transformation and optimization work;finally,the above-mentioned improved results are verified by relevant comparative experiments Effectiveness.Finally,an experimental platform is built to verify the effectiveness of the improved visual odometry,which is mainly verified by two ways:(1)the performance of the whole system before and after the improvement is tested by using the open data set,and the accuracy of the space motion trajectory generated by the system is evaluated by introducing the relevant quantitative standards;(2)two data sets with different degrees of complexity are selected,through Local and global 3D point cloud maps are constructed to verify the robustness and accuracy of the improved system.