An Adaptive Multiscale Scheme for Real-Time Motion Field Estimation

The problem considered in this work is that of estimating the motion field (i.e. the projection of the velocity field onto the image plane) from a temporal sequence of images. Generic images contain different objects with diverse spatial frequencies and motion amplitudes. To deal with this complex environment in a fast and effective way, biological visual systems use parallel processing, visual channels at different resolutions and adaptive mechanisms. In this paper a new adaptive multiscale scheme is proposed, in which the spatial discretization scale is based on a local estimate of the errors involved. Considering the constraints for real-time operation, flexibility and portability, the scheme can be implemented on MIMD parallel computers with medium size grains with high efficiency. Tests with ray-traced and video-acquired images for different motion ranges show that this method produces a better estimation with respect to the homogeneous (no Gadap t ive) mult iscale met hod.