| A system that handles two computer-vision tasks related to motion analysis is presented. Using a pair of grey-level images as input, the system computes: (1) an approximation of the 2-D visual velocity field of a temporally changing scene, and (2) detects motion boundaries.; The proposed algorithm uses motion blur in a novel way. It exploits a new image acquisition method that acquires two images simultaneously, in which one image is a temporal extension of the other. This avoids the need to estimate temporal derivatives, essential for algorithms that are based on the gradient constraint equation. Not being based on image sequences, the algorithm also avoids the need to follow image features from one frame to the next. This elegantly avoids the correspondence problem that is associated with almost all visual-motion estimation techniques. Through experiments, it is demonstrated that appropriate analysis of motion blur results in an unbiased estimator for the orthogonal component of the 2-D image velocity vector. It is also shown that the algorithm is able to segment objects in motion without the need to solve for the optical flow field.; The algorithm exploits integral operations in almost all of its computational sub-processes to reduce the effects of noise. All computations are performed in the spatial domain for ease of future implementation in analog VLSI. The success of the method comes in large part from its ability to integrate effectively motion blur as a raw source of information with a simple, local non-iterative procedure to find displacements of moving edges. |
