A High-Precision Calibration Method for Stereo Vision System

Stereo vision plays an important role in planetary exploration, for it can percept and measure the 3-D information of the unstructured environment in a passive manner (Goldberg et al., 2002; Olson et al., 2003; Xiong et al., 2001). It can provide consultant support for robotics control and decision-making. So it is applied in the field of rover navigation, real-time hazard avoidance, path programming and terrain modelling. In some cases, one stereo-vision system must accomplish both hazard detection and accurate localization with short baseline, i.e. 100-200mm in length. This seems to be a little ambivalent, for hazard detection needs wide view field, while accurate localization is on the contrary. Reconstruction precision is inverse proportion to focal length if the baseline is fixed. So researchers have to first select a compatible view angle, which guarantees the task workspace is within the view field. Then they must refine their camera calibration method in order to satisfy the accuracy requirement of rover localization, navigation and task operation. In order to satisfy these requirements, wide angle lens is usually used. Lens distortion may reduce the precision of localization. So distortion parameter calibration plays an important part in such case. Moreover, calibration accuracy may also affect the complexity of the matching process. Tsai (R, Y, Tsai, 1987) proposed a method, in which a distorted parameter is used to describe the radial distortion of the lens. A five-parameter model is exploited to characterize several kinds of lens distortion (Yunde et al., 2000). A more complicated model, CAHVORE, is introduced (Gennery, 2001). Calibration becomes a nonlinear process if lens distortion is introduced. Usually camera calibration needs two steps. The first step generates an approximate solution using a linear technique, while the second step refines the linear solution using a nonlinear iterative procedure. The approximate solutions provided by the linear techniques must be good enough for the subsequent nonlinear technique to correctly converge. After the initial value has been obtained, the precision of the final result and convergence speed depends closely on optimization algorithm. Most of existing nonlinear methods minimize the geometric cost function using variants of conventional optimization techniques like gradient-descent, conjugate gradient descent Newton or LevenbergMarquardt (LM) method. Therefore there are some problems in these circumstances. First, it is the commonly used cost function, reprojection error, which minimizes the distance