Direct dynamic motion vision

A method is presented for estimating the structure of a scene and the motion of an observer relative to the scene from a sequence of images. Two key features distinguish this method from previous solutions to this problem: no computation of optical flow is required, which leads to considerable speedup, and a Kalman filtering algorithm takes advantage of the entire sequence and leads to noise reduction. Dense depth is estimated by the Kalman filter and is obtained from motion by a least-squares method. No assumptions about surface structure or motion are made. Experimental results on real images are presented.<>