Proceedings PaperModel for shape and motion perception
|Format||Member Price||Non-Member Price|
In this paper we describe an efficient system for recovering the 3-D motion and structure of visual systems from an evolving image sequence. The technique utilizes the image flow velocities in order to recover the 3-D parameters. We develop a real-time algorithm for recovering the 2-D flow vectors on the image plane with sub-pixel accuracy. The flow estimates are then examined for the possibility of errors, mistakes, and uncertainties in the visual system. Uncertainty models are developed for the sensor and for the image processing techniques being used. Further filtering and a rejection mechanism are then developed to discard unrealistic flow estimates. The 2-D flow models are then converted into 3-D uncertainty models for motion and structure. We further develop an algorithm which iteratively improves the 3-D solution given two successive image frames and then discuss a multiframe algorithm that improves the solution progressively by using a large number of image frames.