Presents a model-based algorithm for estimating motion frommonocular image sequences. The authors first present a two-view motionalgorithm and then extend it to multiple views. The two-view algorithmrequires generally 6 pairs of point correspondences to give uniquesolution of the motion parameters. However, when the used points lie ona Maybank quadric, the algorithm requires 7 pairs of pointcorrespondences to give double solutions. Object-centered motionrepresentations and a motion model of constant acceleration are used toestimate motion parameters from long image sequences. The algorithmguarantees globally optimal solution. Since the algorithm does notinvolve structure parameters, it contains the least number of unknownsand is hence more efficient and robust than the existing ones.Experimental results with real image data are presented. The same methodcan be applied to solve for motions described by second or higher ordersof polynomials
展开▼
