Research On Visual Odometry Technology Based On Dynamic Scene

With the rapid development of intelligent robots and unmanned areas,machine vision has been widely used.In the field of robots,visual navigation and positioning technology is a key technology.With images as input information,the visual system that calculates the position information of the carrier is often called visual odometry.The visual odometer framework is based on a static scene and is easily affected by dynamic features.When there are many dynamic features in an image,accurate positioning information cannot be output.This paper studies two algorithms filtering in dynamic feature points based on dynamic scenes.It is applied in visual odometry to improve accuracy of real-time pose resolutionIn this paper,the principle of visual odometry positioning based on stereo CCD camera is studied,and the position and attitude solving equations are deduced.Then,the process and formula of the stereo visual solution poses are deduced,including camera calibration,threedimensional information reconstruction,image feature matching,motion estimation model and global pose updating equation.Aiming at the problem of frame dropping in the system due to insufficient feature points in some specific scenes,an adaptive threshold-based ORB feature extraction method is proposed.The ZED binocular camera is also used for the experimental verification and analysis of key technologies proposed in this paper.In order to overcome the errors caused by too many dynamic features in the scene,a combined filtering scheme based on RANSAC algorithm and a support vector machine based target recognition algorithm is proposed.Finally,a real-time visual localization system based on dynamic scenes was established.The KITTI data was used for off-line processing and the experiments were conducted on urban roads and highways.In a variety of scenarios,using ZED cameras for real-time positioning to verify the robustness and accuracy of the stereo visual system,and analyzing the experimental errors in various aspects.This study is of important reference value for subsequent optimization.