Research On Stereo Vision SLAM Based On Manhattan-World Assumption

Vision-based simultaneous localization and mapping(SLAM)is an important method for robots to perceive the environment.Although feature-based visual SLAM can estimate the position and pose of the camera well,it will generate large accumulative error over time.Most man-made environments contain the Manhattan world coordinate system.Because the Manhattan world coordinate system does not change with the motion of the camera,it can be used to globally constrain the orientation estimation of the SLAM system.In view of this,this paper proposes a stereo vision SLAM system based on the Manhattan world assumption,which has certain engineering application value for the localization and mapping of the robots working indoors.The main research work is as follows:In the stereo vision SLAM system,the random sampling consensus algorithm and the branch and bound method are combined to establish the Manhattan world coordinate system.A globally consistent Manhattan world coordinate system is established while satisfying real-time performance.And use the global consistency of the Manhattan world coordinate system in the local map to continuously optimize the Manhattan world coordinate system.In order to utilize the scene structure information as much as possible in complex environments,when the total number of lines associated with the Manhattan world coordinate system is less than the threshold,a sub-Manhattan world coordinate system with the same z-axis as the Manhattan world coordinate system is established.In order to reduce the accumulative error of rotation estimation during pose estimation,the rotation estimation of camera is optimized using the connection between the Manhattan world coordinate system and the structure lines.The Manhattan world coordinate system is established in the camera coordinate system of each key-frame,and the global consistency of the Manhattan world coordinate system is used to further reduce the rotation accumulative error of key-frames.In order to reduce the cumulative error of key-frames’ pose estimation,the Manhattan world coordinate system in the camera coordinate system of key-frames in the local map is associated to optimize the key-frames rotation estimation.Different from visual SLAM that only uses structural lines as landmarks,this paper uses common lines and structural lines as landmarks to improve the robustness of the SLAM system.And rationally parameterizes landmarks of structural lines to improve the optimization efficiency of the SLAM system.Tracking and localization experiments are performed on indoor scene datasets,outdoor scene datasets and real scenes.The experimental results in the indoor scene datasets prove that the rotation optimization algorithm based on the Manhattan world assumption proposed in this paper can significantly improve the rotation estimation accuracy of visual SLAM without affecting the translation estimation accuracy of visual SLAM.The experimental results in the indoor scene datasets prove that the rotation optimization algorithm based on the Manhattan world assumption proposed in this paper can significantly improve the rotation estimation accuracy of visual SLAM.The experimental results in the outdoor scene datasets prove that the reasonable use of the Manhattan world assumption and the full use of line-landmarks make the visual SLAM system proposed in this paper have good robustness.Experiments in real scenes demonstrate that the proposed visual SLAM can robustly and accurately estimate the camera pose.