Parallel computations of optic flow in early image processing

Current calculations of optic flow from the motion of edge features in an image sequence are described, and used to illustrate the implementation of low-level image processing calculations on an SIMD (single instruction multiple data) machine. Most of the computational effort in the optic flow calculations is required for carrying out large, multidimensional, spatiotemporal convolutions over the image data, which may be carried out efficiently in parallel by mapping the image pixels into the processor array. Examples are given, showing that computation times consistent with video data input rates can easily be achieved. Later stages of the optic flow calculation, after moving edges features have been detected, are also carried out in parallel, although the computations are no longer completely uniform over the image data arrays. A VLSI processor array, embodying many of the features required for low-level image processing calculations, is briefly described.