This talk addresses the problem of recovering both the camera motion and scene structure given a sequence of images obtained from a single camera moving through a rigid environment. I will review the model-based brightness constraints, a method for computing the camera motion and dense scene structure from three views. I will show results laboratory and outdoor environments.

We can apply the same framework to the analysis of image sequences obtained from a moving stereo rig. The motion of the stereo rig can be computed directly from the image derivatives using linear equations. Correspondences are not required. The motion estimates between stereo pairs enable us to combine depth maps from all the pairs in the sequence to form an extended scene reconstruction.

For accurate ego motion estimation a wide field of view is required. For wide angle lenses, non-linear lens distortion can be a significant factor and affect the accuracy of motion and structure estimates. I will present a method for determining the lens distortion parameters using geometric constraints on a set of point correspondences in multiple views (CVPR97). There is no need to know the 3D location of the points.