Research On SLAM Algorithm Based On Nonlinear Optimization And Depth Map Comparison

With the rapid development of computer vision,autonomous driving and mobile robotics are becoming more and more mature,and Simultaneous Localization and Mapping(SLAM)technology is one of them used to solve the problem of localizing the location of mobile robots in an unknown environment and building a map of the environment they are in.SLAM algorithms can be mainly classified into visual SLAM and laser SLAM by the category of sensors used in the front end.Visual SLAM algorithms can be competent for indoor textured and static scenes,but the cameras are easily affected by ambient light leading to poor robustness of the algorithms.The laser SLAM algorithm can run outdoors with less texture,is not susceptible to environmental influences,and can be prone to insufficient information and incorrect information with single sensor,and the SLAM system will get misestimation results or even failure.In order to enhance the robustness of the SLAM system and to solve the problem of bias in the estimation of the poses caused by fewer textures and similar environmental features,this thesis studies the robust multi-sensor fusion SLAM system,which uses the multi-sensor fusion SLAM algorithm and the dynamic obstacle filtering algorithm to optimize the localization and map building problems in SLAM technology respectively,with the following main contents and completed works:The multi-sensor fusion SLAM algorithm based on nonlinear optimization is designed to address the problem of offset in positional estimation caused by environmental factors.The algorithm is divided into two subsystems,laser and vision,and the positional estimation process of both subsystems uses nonlinear optimization.The laser subsystem is a LIDAR and IMU tightly coupled system,which uses the distances between the corner and plane points extracted from the laser frame and the local map as residuals for the positional optimization,and introduces the weight value of laser point curvature in the residuals.The vision subsystem is a tightly coupled system of vision camera and IMU,which adds constraints from laser frame matching to the objective function of the factor map.The results of the control experiments show that the algorithm of this thesis significantly improves the localization accuracy.The dynamic obstacle filtering algorithm based on depth map comparison is designed for the problem of dynamic obstacles interfering with map reuse.The method uses depth map comparison to determine dynamic points based on dynamic obstacle filtering based on viewpoint visibility.The laser point coordinates are converted using spherical coordinates,and the dynamic points are determined by whether the laser point optical path crosses another laser point,and the static points are recovered from the set of dynamic points by cycling the depth map comparison with different resolutions.The experimental results show that the algorithm effectively filters out dynamic obstacles in the map.A multi-sensor fusion SLAM experimental platform based on a four-wheel differential chassis was designed and implemented.Firstly,sensor selection,hardware and software adaptation and driver installation were carried out to build a mobile robot experimental platform based on ROS framework.Subsequently,the multi-sensor fusion SLAM algorithm and dynamic obstacle filtering algorithm were experimented on the experimental platform in a real vehicle,and the experimental results proved the robustness and effectiveness of the algorithm by using public data sets together.