Direction-selective Lters for Egomotion Estimation

The dimensionality of visual motion analysis can be reduced by analyzing directional components (projections) of ow elds. In contrast to vector elds, these lower-dimensional spaces exhibit two simple geometric properties which are invariant to the scene structure and depend only on the camera motion. Using these properties, structure and motion can be separated. The approach is closely related to the concept of direction-selective processing of visual motion elds, for which biological evidence has been discovered in mammalian visual systems. We use a recursive observer model where a collection of lters tuned to speciic ow directions provide the directional motion parameters. The original motion parameter vector can be computed (if necessary) by combining the directional parameters as vector components. The model is applicable to general camera motion and to large camera eld of view (FOV) and does not require point correspondence. In addition to the recursive model, the temporal integration of instantaneous measurements is extended to image sequences using tracking, which facilitates reconstruction of the camera motion trajectory, even if the velocity of the motion changes. The approach is highly scalable and eeciently parallelizable. We demonstrate it on long image sequences.