Research On RGBD SLAM Algorithm For Indoor Low-Texture Scene

Indoor robot localization and navigation is the key technology to achieve autonomous motion control for robots.There are still some challenging scenarios for camera-based indoor localization algorithms.For example,the existing algorithms suffer from poor accuracy and poor robustness when the texture is low or the illumination is poor.To address the problems of SLAM systems in low-texture scenarios,changes have been made based on VINS framework in this work.By combined a feature tracking algorithm with depth gradient and IMU data,a more robust and highly accurate feature tracking and pose estimation algorithm for texture-sparse scenarios is proposed.The back-end is simultaneously constraining visual observation and inertial measurement to estimate the time difference(with camera and IMU).The new back-end will further improve the localization accuracy.The algorithm is finally applied to the mapping and navigation process of an indoor quadruped robot.The main contributions of this work are:In the front-end part of the SLAM algorithm:The optical flow estimation algorithm is introduced and the possibility of only using the depth map as the input for optical flow estimation is discussed.A novel optical flow estimation algorithm is proposed based on the traditional optical flow estimation algorithm combined with the depth map.The biggest highlight of the proposed algorithm is to extend the 2D optical flow estimation to 3D space.Meanwhile,the inaccuracy of the pure depth map for the direct estimation of optical flow is proved.Secondly,after proposing the new optical flow estimation algorithm,an improved direct method algorithm is proposed by analogy with the gray-scale map-based optical flow estimation algorithm and the direct normal pose estimation algorithm.The biggest highlight of the proposed direct method pose estimation algorithm combined with depth map is the use of depth gradient information for searching,while the current direct method pose estimation algorithm combined with depth only uses the depth information for constructing projection relations.In the back-end part of the SLAM algorithm:The effects of adding temporal offsets to visual observations and IMU measurements,are discussed respectively.The camera and IMU observation models are re-modeled.Then,based on the back-end sliding window optimization method of VINS,a least-squares problem is constructed to optimize the time offsets from both visual and IMU aspects simultaneously.Finally,the improved effect of the algorithm is evaluated in blocks,and a quadruped robot experiment platform is built for engineering implementation.In the experiment,when the depth image quality is good,the pose estimation result of the front-end algorithm proposed in this paper has a better error of 13% of the original algorithm.The time offset convergence time reduction in the improved back-end algorithm improves the pose optimization accuracy by 4%.Moreover,the three functions of real-time pose estimation,cloud map building,and navigation are realized on the quadruped robot.