Research On Global Self-Localization Of Indoor Mobile Robot Based On RGB-D Camera

Localization is the process in which mobile robots determine their position in the scene by sensor information.Accurate and robust self-localization is the basic condition for mobile robots to complete navigation tasks.Global localization requires searching in the feasible pose state space in the map.The calculation of positioning process is enormous.In the process of positioning,many similar scenarios,local scene changes,and dependence on prior information acquisition path are easy to cause positioning failure.RGB-D camera has low cost and can obtain both color and depth information.Indoor navigation and positioning based on RGB-D camera has become a hot research topic in the field of mobile robots.However,due to the limitations of sensor performance,the complexity of three-dimensional visual information processing and the diversity of application scenarios,there are still many difficulties and challenges in mobile robot localization based on RGB-D camera.According to the characteristics of RGB-D camera,this paper studies the key technology of three-dimensional vision self-localization for indoor mobile robots,in order to improve the efficiency,accuracy and robustness of localization.The main contributions of this paper are as follows:Firstly,image retrieval based on the Bag of Visual Word model is used to determine the global position and posture of mobile robots,and an active search and location strategy is proposed,which can actively acquire more visual information after the failure of localization from the current observed information and improve the robustness of localization.Based on the correlation between image sequences captured in the motion process,the optimal global positioning estimation is determined by statistical analysis,which improves the efficiency and accuracy of global positioning,and provides a good initial positioning result for the precise positioning and navigation of mobile robots.Secondly,three-dimensional point cloud registration is the key technology for mobile robot location through three-dimensional point cloud.The real-time localization of robot requires more efficient point cloud registration technology.The point cloud data obtained from RGB-D camera have less overlap area due to the small field of view of the camera,and the noise of the depth camera is large and the accuracy is low,which requires that the point cloud registration has stronger robustness.In this paper,an improved point cloud registration algorithm is proposed,which combines 3D-NDT point cloud registration algorithm with point cloud registration idea based on feature matching.The real-time performance of point cloud registration is significantly improved by feature matching of three-dimensional raster data transformed by NDT algorithm,and the registration between three-dimensional point cloud data obtained by depth camera has stronger robustness.Finally,the classical Monte Carlo localization algorithm,which is widely used at present,still has some problems,such as poor global positioning ability,insensitive detection to the problem of kidnapping robots,and long recovery time.According to the characteristics of RGB-D camera,this paper combine the global positioning method based on bag of visual word and the improved point cloud registration algorithm,based on particle filter positioning framework,designed and implemented an efficient and robust Monte Carlo positioning system based on RGB-D information.For the kidnapping problem,the paper proposed a detection mechanism which based the variation of particle weights and the constraints of real-time observation image sequence continuity.Experiments show that compared with the classical Monte Carlo positioning method,the proposed positioning system has higher robustness and positioning accuracy,and can detect kidnapping problem better and recover positioning quickly.