Research On Vision-based Traversable Region Recognition For Mobile Robots In Outdoor Environment

Fast and accurate traversable region recognition is of vital significance toautonomous navigation for mobile robots. Vision, as a key function and means forhuman’s observation and cognition, is able to provide the richest information. Thisdissertation performs research on vision-based traversable region recognition for mobilerobots in outdoor environment, and lays the foundation for autonomous navigation ofoutdoor mobile robot by designing and realizing relative algorithms.Firstly, binocular stereo vision is chosen according to the characteristics ofdifferent vision systems. Moreover, perspective projection model of the camera isintroduced, including pinhole perspective model, common-used camera coordinates andthe transformation among different coordinates. Also, basic issues of binocular stereovision are analyzed from four aspects: camera calibration, image rectification, stereomatch and triangulation. And ELAS stereo matching algorithm is selected to fulfill thefast and robust demands of mobile robots.Secondly, U-V disparity theory and construction methods are detailed. Also, themodel of binocular stereo vision system is established under reasonable assumption andprojections of the planes from the world coordinate to the U-V disparity image areintroduced. Additionally, traversable regions are defined according to the U-V disparitytheory. And, an improved algorithm for constructing V-disparity image is proposed andused for traversable region recognition. In this algorithm, V-disparity image isestablished according to non-obstacle disparity image, and it can deal with the problemthat when the ground takes only a small part of the image or dominant obstacles occurin the image, the ground correlation line is difficult to be detected by using traditionalmethods. The thorough experiments are performed by using the mobile robot platformconstructed by ourselves in various outdoor environments to validate the effectivenessand efficiency of the proposed algorithm. Statistics shows that the successful rate oftraversable region recognition can reach92.3%by utilizing our algorithm.Lastly, based on the traversable region information, a grid map is constructed forautonomous navigation of mobile robots. By calculating the percentage of pixelsbelonging to traversable region in each grid, the grid is classified to be safe one orun-safe one. Also, classical A*algorithm is applied to realize path planning between thedefined origin point and the target point in the grid map. The experimental resultsvalidate the efficiency of the constructing algorithm of the grid map and path planning.