Real-time3D Terrain Reconstruction System In Lunar Rover Simulation

With the development of the Chinese lunar exploration program, the lunar rover is ex-pected to land on the moon in2013.The rover will meet the test of complex lunar environ-ment and the most significant performance of the rover is to navigate itself without the help of humankind. As the basis of automatic navigation is obstacle detection and terrain recon-struction, this thesis proposed a real-time3D terrain reconstruction system based on stereo vision, which can help the rover to recognize obstacle around and arrive at the destination successfully.The proposed system based on stereo vision, basically including the following steps: camera calibration, stereo matching, depth recovery, point cloud refinement and terrain rendering, is tested on the test field of Chinese lunar rover after tested on the lunar rover simulation virtual environment.First of all, in order to improve the real-time performance of the system, the self cali-bration method is adopted. After comparing the current algorithm of camera self-calibration, a new theory of linear camera self-calibration presented by Wu Fu-Chao and Hu Zhan-Yi is chosen to calibrate the camera, meanwhile, an improved normalized8-point algorithm is proposed to calculate the fundamental matrix.Secondly, in order to reconstruct the terrain more quickly, the area based stereo match-ing method which can produce dense disparity map is introduced. In detail, an improved adaptive binary widow(ABW) based matching algorithm is proposed in this paper, which firstly removes the unreliable pixels according to the uniqueness constraints of stereo matching in computer vision when the initial disparity map is completely computed with the ABW algorithm, then determines the disparity value of unreliable pixels by interpolation using the method of weighted mean, lastly smoothes the disparity map and removing the noise of the graph with the application of median filter method. Finally, with the camera parameters and disparity obtained above, the depth of the each pixel can be calculated and the point cloud which forms the base data of the terrain digital elevation model (DEM) is then available. The original DEM data is refined with the method of point-wise interpolated algorithm before meshed using split-and-merge algorithm. Fur-thermore the terrain mesh is rendered and the3D terrain is reconstructed. As the system perform very well in virtual environment, further more it is tested on the test field of Chi-nese lunar rover, the result is also satisfactory.