Scene Recognition And Modeling For Mobile Robots In Outdoor Environment

With the increasing in types and application of mobile robot, the application domain expands from structured indoor environment to quasi-structured or unstructured outdoor environment. As the basis of mobile robot’s localization, navigation and exploration, scene recognition and modeling for outdoor environment become one of the hot topics in mobile robot. This dissertation starts form the problem of stair detection and parameter estimation in3D scene, and then conducts an in-depth study on classification and terrain modeling for typical outdoor scene, so as to provide technical support for the mobile robot’s self-adaptation to outdoor environment.In mobile robot’s autonomous navigation, stairs can be seen as obstacles, alternative pathway and also important marks in localization and navigation. Targeting the structural diversity of stairs and distribution uncertainty of3D laser point cloud, an adaptive stair detection and parameter estimation method based on stair topology model and fuzzy set theory is proposed. According to the topological relations of staircase profile, a stair edge detection method based on Angle Histogram Algorithm is proposed to improve the estimation accuracy of the stair edge position. Adopting the in-level line extraction and cross-level line linking strategy, the candidate stair edge line set is constructed effectively. The cascade probability of candidate stair edges between levels is estimated by fuzzy transform and adaptive fuzzy reasoning. Global optimum candidate edge line combination is searched by Simulated Annealing Algorithm, so as to construct the3D stair model effectively.The vehicle vibration and lighting conditions change will significantly affect the quality of vision image. By using Gabor Filter for image enhancement, the stair edge detection is then achieved by adopting Canny Operator. A local fusion phase grouping method is proposed to enhance the robustness of edge line extraction in low quality image. The offset parameter for robot motor control is obtained by linking and filtering of the edge lines. By the feature fusing of monocular vision image and3D laser point cloud data, the reliability of stair detection and parameter estimation is then enhanced.For the mobile robot with certain ability of obstacle negotiation, terrain classification and terrain modeling are an important bases for the discrimination of environment passability. According to the diversity of outdoor terrain complexity, a hierarchical terrain classification and geometric modeling method is proposed for mobile robots’scene recognition and motion planning. Based on the terrain representation of layered elevation map, rapid discrimination of passable area and obstacle area on flat road surface can be realized. A senior feature representation of the inherent ambiguity of natural scene is produced by fuzzy inference, then terrain classification of3D point cloud is achieved according to the principle of maximum entropy, thereby the discrimination veracity of passability in outdoor scene is enhanced. Based on this, using semantic segmentation algorithm of terrain fragment, the modeling to typical terrain structure in outdoor scene is achieved.The effectiveness and practicality of the methods proposed in this dissertation is verified with the experiments on the unmanned ground vehicle (UGV) developed by our research group and the leg-wheel robot developed by Shenyang Institute of Automation.