Research On Visual Based 3D Scene Perception Techniques For The Mobile Robot

Mobile robot perception of three-dimensional scenes is a hot topic in the robotics research recent years, act as the basis for autonomous robot navigation and perform tasks in unknown environments. In this paper, we focus on the study of three-dimensional scenes perception technology based on visual informa tion. Mobile robot perception in three-dimensional scene includes three aspects: the geometry of three-dimensional scene perception, the mobile robot pose perception and the scene target perception. Specifically, this paper relies visual sensor data, calculated based on the binocular stereo vision depth map from fast, real-time mobile robot pose estimation, 3D model reconstruction and 3D object recognition four aspects of the practical application of engineering to carry out theoretical and applied research.Binocular stereo matching, we designed a low- level feature description operator based on the Sobel filter response, matching left and right images, we choose the stabilize matched point and their disparity as a set of supporting points, and the entire image plane is divided into different triangle planars, as each the triangular plane vertex's disparity known and the priori distribution function of disparity on the the region, estimating the other triangular plane pixel disparity. Combined with the powerful GPU parallel computing power,the algorithm achieve a rapid depth map estimation with dense scene.Mobile robot three-dimensional scene reconstruction, this paper uses a method based on surface stitching ICP. Firstly, using GPU parallel computing, the surface data be down-sampling twice into three scales. Then, combined with the fact of small relative movement between the continuous surfaces, simplify the process of solving rigid transformation matrix between surfaces. Achieved the surface alignment and estimating the mobile robot pose information at same time. Finally, we use Marching C ubes algorithm to extract the surface from the volume of the space and establish a dense mesh surface.Three-dimensional object recognition in the real scenes, based on the three-dimensional model reconstruction, the target model will be uniform sampled in three-dimensional space. For each partial view of the model, calculates its the shape index map and SIFT features as a local feature description, and build a model feature library. Finally, match the local feature of scene and the model, achieved the object be effectively recognized under partial occlusion situation.