Motion-Model-Based Boundary Extraction and a Real-Time Implementation

摘要

Motion boundary extraction and optical flow computation are two subproblems of the motion recovery problem that cannot be solved independently of one another. These two problems have been treated separately. A popular recent approach uses an iterative scheme that consists of motion boundary extraction and optical flow computation components and refines each result through iteration. We present a local, noniterative algorithm that simultaneously extracts motion boundaries and computes optical flow. This is achieved by modeling 3-D Hermite polynomial decompositions of image sequences representing the perspective projection of 3-D general motion. Local model parameters are used to determine whether motion should be estimated or motion boundaries should be extracted at the neighborhood. A definite advantage of this noniterative algorithm is its efficiency. It is demonstrated by a real-time implementation and supporting experimental results.