A stream processing architecture for real-time implementation of perspective spatial transformations

A system is presented to perform perspective spatial transformations at video-rate using a mapping function which can vary at frame rate. The mapping is implemented by two successive applications of a one-dimensional algorithm operating in scan-line order using a framestore to store an intermediate image. A novel one-dimensional filtering algorithm provides a continuous mapping from source to destination co-ordinates, preventing aliasing. These algorithms are implemented using a stream processing architecture, an efficient mechanism for video-rate operations. Emphasis is placed on the design of a novel stream processing architecture which, coupled with an efficient pipelined co-ordinate generation system, provides single-cycle per output pixel performance