PRECISION ANALYSIS OF VISUAL ODOMETRY BASED ON DISPARITY CHANGING
Keywords: Visual Odometry, Disparity Changing, Mobile Mapping System, Six Degrees Of Freedom, Pose Estimation
Abstract. This thesis aims to analyze the precision of Position and orientation of cameras on Mobile Mapping System (MMS) determined by disparity based VO (DBVO). Dual forwards taken cameras on MMS are applied to obtain a sequence of stereo pairs. The Interior Orientation Parameters (IOPs) and Relative Orientation Parameters (ROPs) are derived in advance. The pose estimation is achieved by DBVO without additional control data. The procedure of DBVO consists of four steps. First up, keypoint detection and matching is conducted to obtain tie points in consecutive images. Then, image rectification is implemented to transform tie points into epipolar image space. Next, parallax equation is applied to estimate the 3D coordinates of interest points in epipolar image 3D space. Since their image points have different disparity in neighboring stereo pairs, the 3D coordinates of interest points in neighboring pairs are different as well. Finally, 3D conformal transformation is employed to derive the transformation parameters between neighboring pairs according to changing of coordinates of interest points. The posteriori STDs are adopted to assess the quality of transformation. Besides, check data of ground trajectory derived by photo triangulation are applied to evaluate the result. The relative errors of horizontal and vertical translations derived by DBVO are 2 % and 3 % in non-viewing direction. However, the translation in viewing direction and three rotation angles derived by DBVO have significant systematic errors about 1 m, 3°, 3° and 10° respectively. The influence of error propagation is not significant according to the chart of error distance ratio. In open area, the trajectory of INS/GPS is similar to ground truth, while the trajectory derived by DBVO has 44 % relative error. In residential district, the trajectory derived by INS/GPS has drift error about 2 m, while the relative error of the trajectory derived by DBVO decreases to 38 %. It is presumed that the systematic error results from 3D coordinates estimated by parallax equation because of poor intersection geometry. It will be proved by adding sideward photographing cameras in the future.