Research On Robust Visual SLAM For Ground Mobile Robots

Because of ground mobile robots can perform many repetitive,heavy or dangerous tasks for human beings,it has grate value of real application in civil,military and scientific research.Simulation Localization and Mapping(SLAM)is the premise for mobile robots to realize intelligent and autonomous motion,so it also has important research significance and practical value.Due to the low cost,low power consumption,small size and abundant information provided by visual sensors,visual SLAM has always been one of the research hotspots in the field of mobile robots.After more than a decade of research,there have been a lot of results on visual SLAM.On the basis of summarizing the research of visual SLAM in recent years,this paper introduces some background knowledge of visual SLAM in detail,and analyses some robustness problems existing in visual SLAM at this stage.Focusing on the robustness of visual SLAM for mobile robots,two factors affecting the robustness of visual SLAM are studied: 1)dynamic scenes,2)wide scenes,and improvement measures are proposed respectively.This research has certain theoretical significance and value for improving robustness and realizing the universality of robots.The main contents of this paper are as follows:1)Research of robust visual SLAM in dynamic scenes.In dynamic scene,incorrect data association caused by moving objects will make the results of SLAM move away from the real value,resulting in poor robustness of most existing SLAM algorithms in dynamic environments.Although the current robust model estimation algorithms,such as Random Sample Consensus(RANSAC),can better filter the dynamic features,when the low-speed motion persists for a long time,the dynamic points may not be filtered because the model fitting error is less than the threshold value.Because the inter-frame difference can detect slow moving objects very well,this paper proposes an improved dynamic SLAM scheme based on motion feature removal by three-frame differential.Firstly,background compensation is applied to the input image sequence,the images from different view point are compensated to the same.Then,the motion region is detected by three-frame differential method.Finally,the matching points falling on the motion region will be eliminated,which can remove the error data association.Experiments were carried out to verify the effectiveness and feasibility of the improvement.2)Research of robust visual SLAM in wide scenes.Because of to the limitation of depth range,Stereo SLAM and RGBD-SLAM have poor robustness in wide scenes,while Monocular SLAM has no limitation of depth range,has better robustness in open scenes.In order to improve the robustness of open scenes,this paper studies the scale recovery algorithm of Monocular SLAM.In this paper,an algorithm for recove the local scale of Monocular SLAM by ground matching is proposed.The reference value of the real motion of the camera can be directly estimated using the planar motion model and the installation height of the camera.The scale parameters can be obtained by calculating the similarity transformation between the results of Monocular SLAM and the reference values.Experiments with ORB-SLAM show that this method is feasible.