Research On Three-Dimensional Planar Surfaces Map Building For Robots

For the presence of a large number of planar features both indoors and outdoors and point features are susceptible to light and other factors,for instance, when light is dim and wall smooth, point features can be vague even invisable. So the method of three-dimensional planar surfaces map building is to solve the problems above.To reduce errors, the closed-loop detection method is proposed based on planar feature matching and optomize the overall map. The main work is as follows:First, create the map platforms in the Ubuntu system. Software platform includes some visual libraries as PCL, OpenCV, and ROS, and hardware platforms including kinect camera and its carrying platform Pioneer second generation.Secondly, bring the point cloud data to plane detection by the region growing method based on PCA to have three-dimensional plane segmentation.First estimate the normal vector by PCA(principal component analysis) to calculate the point cloud with the same direction of the normal vector quasi local plane normal vector angle and the distance to the origin of coordinates satisfy the constraint conditions,merge the local planes through region growing method,then build the 3D plane segmentation of point cloud and extract planar characteristics.Then introduced a complete plane patch on the basis of the segmentation, then chose three pairs of pairwise non-parallel planar patch to replace image frames in registeration by the information of image frame contained in the complete planar patch. According to the three constraints, find the correspondence of a complete planar patch in two frmes and calculate the pose relations among frames (rotation matrix and translation matrix) to achieve image registration. However, due to the accumulated errors in image registration which may drift the map, there is need for further optimization.To solve the problems above, this paper proposes to find neighborhood of the current frame by using the Euclidean distance firstly, and match the neighborhood frame based on planar features in a innotive way.Save the previous planar features of frames, then extract the planar features of the later ones, and match the planar features of all. The match is successful when the three threshold conditions are met.Detect the closed-loop and optimize the map by the method of gradient descent (SGD, Stochastic Gradient Descent).The experiment shows that the map display has been improved significantly.