Terrain reconstruction of lunar surface based on binocular fisheye camera

3D scene reconstruction of binocular fisheye obstacle camera for lunar rover is challenging. This paper presents a large field of view (FOV) scene reconstruction method based on fisheye lens. To overcome the problems of the large imaging distortion of large FOV fisheye lens and the existence of information loss when converting to perspective image, we studied the calibration method of spherical model and spherical stereo vision in-depth. Based on unified spherical model, we further propose the generalized unified model and design a complete calibration method, which improves the calibration precision of fisheye wide-angle camera. Large FOV stereo vision method proposed in this paper uses spherical imaging model to describe the stereo camera, and defines the concept of depth reconstruction and the corresponding disparity for spherical model. In order to more effectively carry out the search of corresponding pixels in the spherical curve, spherical model is expanded into latitude and longitude image using conformal projection transformation. Then according to the characteristics of the latitude and longitude image, Semi-Global Matching algorithm is used for solving spherical disparity. Finally, the 3D data of continuous multi frame reconstruction is registered and fused to have a wider range of scene information. The experimental results from simulative lunar surface environment show that, our algorithm can calculate near 180° of the disparity result of the front scene, and recover the 3D information of a wider range of scene compared with the traditional methods. At the same time, the accuracy for the close-up objects which wide FOV lens aim at is fairly high.