Research On Position Algorithm Of Visual Odometry Based On Binocular Camera

Visual navigation is an emerging field of navigation technology.With the development of artificial intelligence-related technologies and people's increasing demand for visual related products and services,visual navigation has broad application prospects.In order to improve the positioning accuracy of the Visual Odometry and reduce the running time of the algorithm,a binocular camera-based Visual Odometry system is constructed in this paper and the algorithms of image feature extraction and tracking,pose estimation and estimation optimization are studied.In order to improve the robustness of the feature extraction algorithm,an o FAST-FREAK feature extraction algorithm based on the combination of o FAST corner detection algorithm and FREAK binary descriptor is designed.This algorithm can extract image features under different changing scenes quickly and effectively.On this basis,a grid-based feature control method is proposed.PROSAC outlier elimination algorithm and cross-checked BF algorithm are utilized to match features.This paper presents an improved DLT pose estimation algorithm.The improved algorithm imposes an overall constraint on the rotation matrix and weights the spatial 3D point information to ensure that the time complexity is low and the accuracy of the original algorithm is improved.By statistically analyzing the mean and median values of the rotation and translation errors of the improved DLT algorithm,the error results are similiar to the mainstream pose estimation algorithms such as EPn P and RPn P,but the computation time is significantly shortened.Based on the pose estimation,a local and global SBA pose optimization scheme based on keyframe unit is designed.The optimization process converts the pose optimization problem into a graph optimization problem and solves it with the open source library G2 O.The Levenberg-Marquardt algorithm is used to solve the least squares problem,and the Cauchy Robust kernel function replaces the standard error function.The pose optimization scheme can effectively suppress the impact of large error terms,ensure the global optimal performance and improve the positioning accuracy.KITTI dataset data and actual data collected by ZED are utilized to verify the effectiveness of the algorithm.Compared with the original algorithm,the proposed algorithm has a low time complexity and the computation time is about 85% of the original algorithm.This proposed algorithm can effectively reduce the positioning peak error,cumulative error and RMS error.