On A Laser Radar-Based 3D Map Building Method Of Outdoor Environments

The traditional mobile robot positioning technology such as GPS global positioning system must rely on external equipment to obtain the target position,and the positioning methods are limited by the positioning scene of the robot,e.g.,the tunnel,underground or road environment blocked by trees,so the GPS positioning results always have serious errors and thus cannot be adapted to the above scenarios.Since the research of robots is gradually developing towards the directions of independent planning path and the intelligent autonomous movement,the future research on robot movement must involve the positioning of multiple scenes.But,traditional high-cost positioning methods can't meet the development of mobile robots.The simultaneous localization and mapping(SLAM)technology,as an emerging research direction,is able to sense environmental information through the sensors equipped on the robot mobile platform and to reduce error of the map by utilizing the coupling relations between the robot pose and the map to amend the drawbacks of traditional positioning methods.The robustness of the laser sensor and the accuracy of detecting the environmental information make it become the preferred sensor for mobile robots creating high-precision outdoor environment and accurate positioning.Based on the above discussions,this thesis investigates the following problems based on the Velodyne laser sensor SLAM technology.First,we consider the problem that the laser sensors product the noise interference data reducing accuracy of map when sensing the external environment.The question is studied from the two aspects of data preprocessing and point cloud registration.This thesis adopts a statistical point cloud filtering algorithm to eliminate irregular data and reduce the influence of 'spike' noise on map.In addition,data preprocessing has both a conversion data format and custom data sparseness in order to meet the real-time requirements of creation map.Next,the matching process of the point cloud smooth the local point cloud data through the normal distributions transform(NDT)algorithm,thereby solving the problem of local data loss or abnormal data caused by unknown factors when the sensor collects the external environment data.Considering the problem of data frame matching low rate and low robustness caused by strong coupling of point cloud data registration in the process of simultaneous positioning and mapping,a kind of point cloud registration methods combining the "inverted tree" point cloud processing sequence and the NDT algorithm is proposed.By changing the matching order of adjacent data frames,the problem of low stability of the mapping system caused by the cumulative error of the point cloud data sequential registration is avoided.This method has the advantage of multi-threaded calling NDT registration algorithm to improve data processing speed.As for the problems that the positioning error generated by the pose accumulation when creating a map in the outdoor large environment,and that the closed-loop scene is difficult to get accurate map,this thesis proposes a closed-loop recognition scheme based on the Euclidean space distance.According to the graph-based simultaneous positioning and mapping idea with global optimization,the pose of the mobile robot at different moments is taken as the node,and the covariance between the poses is used as the constraint condition of the edge in order to construct the graph structure.By continuous iterative optimization,it is detected whether there is a closed loop,and the accumulated error is dispersed to different nodes,thereby reducing the overall error and producing a globally consistent map.In view of the above problems and proposed solutions,this thesis based on the intelligent factory laboratory of Northeastern University and campus environment has extensively and thoroughly tested and verified the problems existing in the process of simultaneous positioning and mapping.