Research On Positioning Technology Of Stereo Visual Odometry For Autonomous Land Vehicles

The development of autonomous land vehicles has been driven by policies,economy,society,technology and other aspects,which has become an important strategic direction for the future development of automobiles all over the world.The autonomous positioning ability of autonomous land vehicles is of great significance in the process of autonomous driving,which is related to the realization of autonomous decision-making,path planning and other important tasks of autonomous land vehicles.Stereo visual odometry is not only an effective choice for autonomous positioning in outdoor environment,but also it can complement other sensors mounted on the autonomous land vehicles.Therefore,stereo visual odometry for autonomous land vehicles is gradually becoming a popular topic in the field of autonomous positioning technology,which is of great significance to the research and development of autonomous land vehicles.Aiming at the characteristics of autonomous land vehicles and their environment,this paper carries out a further research on the algorithm of stereo visual odometry for autonomous land vehicles.By the way of studying the working principle and mathematical models of stereo visual odometry,this paper selects the main modules reasonably on the basis of the feature-based method’s framework.The basic stereo visual odometry model is obtained through modeling and improvement of these main modules.In addition,the simulation test and analysis of the model lay a foundation for the subsequent research and improvement of stereo visual odometry.Considering that autonomous land vehicles usually travel in very complicated dynamic scene and that visual odometry is usually researched on the basis of static scene assumptions.In this paper,a data association algorithm for dynamic scene is proposed.Firstly,based on image semantic segmentation,a priori semantic information is obtained through SegNet,which contributes to eliminating the dynamic objects in the scene.Then,based on the geometric characteristics of feature points,the mismatches in dynamic scene are eliminated through an improved RANSAC and Delaunay triangulation algorithm.Finally,the data association module of stereo visual odometry is improved by combining image semantic information and geometric characteristics of feature points,and the effectiveness of the algorithm is verified by simulation experiments.The above algorithm can deal with dynamic scenes robustly and effectively eliminate mismatches in dynamic scene.Considering that the camera is constrained by the mounting platform and visual odometry usually estimates the unconstrained 6-DOF motion of the camera directly.In this paper,a pose estimation algorithm for autonomous land vehicles is proposed.Firstly,the pose estimation model of feature-based method is deduced by considering vehicle kinematics constraints and solved by separate estimation and refinement of the pose.Secondly,combining with semantic information,the pixels of the road sign are extracted to estimate the pose on the basis of pose estimation model of direct method.Finally,the pose estimation module of stereo visual odometry is improved by fusing these two pose estimation models,and the effectiveness of the algorithm is verified by simulation experiments.The algorithm can robustly deal with the scene without enough features,and make the result of vehicle pose estimation more accurately.Considering the errors of stereo visual odometry system in the working process,this paper improves the local optimization module of stereo visual odometry system in view of the depth error of stereo vision and the pose error of stereo visual odometry system.Firstly,based on depth filter,the depth of feature points is optimized by inverse depth parameterization technique.Then,based on bundle adjustment and pose graph optimization,the pose results of stereo visual odometry system are optimized by introducing the sliding window strategy.In addition,this paper tests these two parts respectively,which verifies the effectiveness of the optimization,and verifies that the local optimization module can effectively improve the accuracy of stereo visual odometry system through open-source datasets experiments.By synthesizing the basic model of the stereo visual odometry system and the improved main modules,the final stereo visual odometry system for autonomous land vehicles in this paper is obtained,including the algorithm framework,hardware configuration,software environment.The effectiveness of the algorithm is tested and analyzed in detail through opensource datasets and off-line vehicle data.The experimental results show that the stereo visual odometry algorithm in this paper can deal with complex dynamic scene,and its pose results are in good agreement with the actual kinematic characteristics of the vehicle.What’s more,it is superior to the state-of-the-art visual odometry in robustness and accuracy.Therefore,the research of this paper could be meaningful to solve the problem of autonomous positioning for autonomous land vehicles when the traditional positioning technology fails,and lays the foundation for the realization of autonomous positioning for autonomous land vehicles.