Indoor Navigation For Robot Based On RGB-D Camera

With the development of robotics,mobile robots are widely used in industrial manufacturing,defense,anti-terrorism,space exploration,education,medical and entertainment and many other areas.In the mobile robot research and application of the most critical basic ability is to be able to independent navigation in the unknown environment,and this core problem is often divided into the unknown environment for the researchers to determine their own target location,complex path planning.As a result of the popularity and development of visual sensors,the use of RGB-D camera to solve the mobile robot navigation problem also shows a lot of advantages.Based on the RGB-D camera,this paper designs an indoor navigation system for mobile robots.The system follows the RGBD-SLAM technology and image localization technology and path planning technology.The system can accurately reconstruct the three-dimensional model of the indoor environment and determine the initial and target position of the robot and complete the robot navigation task.At the same time,based on the characteristics and difficulties of SLAM technology and image-based location technology,we have studied several aspects:Firstly,a real-time robot navigation system based on RGB-D camera is constructed.We have studied and analyzed the algorithms and principles of image feature extraction,feature matching,robot pose estimation,loopback detection,graph optimization,point cloud reconstruction,active search,path planning in the system.On the basis of the existing,a complete system from the color and depth images to the scene point cloud model and from the initial target color images to the robot position is completed.Secondly,based on RGB-D data SLAM technology,the depth and color images captured by the traveling robot are constructed by the visual odometer,loopback detection and global optimization to construct the scene point cloud model.Considering the accuracy and real-time requirements of the point cloud model in real-time navigation system,a point cloud construction method based on voxel filtering and statistical filtering is proposed,which can effectively reduce the ghosting problem while effectively reducing the number of point clouds,resulting in a real-time robot interior scene model.Thirdly,in the image-based relocation technique,when the 2D points in the image are correlated with the 3D points in the point cloud model,there is a problem that the 2D-3D matching speed is slow and the 3D-2D matching efficiency is low.Therefore,we propose an active method combining 2D-3D and 3D-2D.After obtaining the matching pair,we can get the initial and the target position of the robot by the robot's pose estimation based on RANSAC.