Direct Generation of Regular-Grid Ground Surface mAP from In-Vehicle Stereo Image Sequences

Abstract

We propose a direct method for incrementally estimating a regular-grid ground surface map from stereo image sequences captured by nearly front-looking stereo cameras, taking illumination changes on all images into consideration. At each frame, we simultaneously estimate a camera motion and vertex heights of the regular mesh, composed of piecewise triangular patches, drawn on a level plane in the ground coordinate system, by minimizing a cost representing the differences of the photo metrically transformed pixel values in homography-related projective triangular patches over three image pairs in a two-frame stereo image sequence. The data term is formulated by the Inverse Compositional trick for high computational efficiency. The main difficulty of the problem formulation lies in the instability of the height estimation for the vertices distant from the cameras. We first develop a stereo ground surface reconstruction method where the stability is effectively improved by the combinational use of three complementary techniques, the use of a smoothness term, update constraint term, and a hierarchical meshing approach. Then we extend the stereo method for incremental ground surface map generation. The validity of the proposed method is demonstrated through experiments using real images.