Estimating 3D vehicle motion in an outdoor scene from monocular and stereo image sequences

The main goal of this research is to test how well existing feature extraction, matching and motion estimation algorithms (with appropriate modification) work on outdoor scenes. For this purpose, a careful calibrated image sequence data base has been created. The data used for the results reported in the paper consists of a sequence of 24 stereo images of an outdoor scene containing a moving truck with stationary background. Two motion estimation methods using feature correspondences were applied in the data: point correspondences over two stereo image pairs and line correspondences over three monocular images. In spite of the large values of the range to baseline ration (10:1) and the range to focal length ration (1000:1), the estimated rotation parameters are reasonably accurate (10-20% errors) in both methods. Although the translation estimates in the monocular method are large, the translation errors in the stereo method are around 1 meter, and are mainly due to image sampling