What is next in real-time computer vision? An overview of tightly-coupled non-homogeneous multiprocessor architectures

The problem of limitations of computer-vision systems caused by the power of computational platforms is addressed by using nonhomogeneous multiprocessors in a tightly-coupled organization. The general form of the architecture is defined as the computational resource graph being homomorphic to the computational task graph. Pragmatic issues such as the skewing of time have to be explicitly addressed and modeled in actual system design. An example of a three-chip set for real-time line buffering, convolution, and image normalization, is presented; this set encompasses modularity of the control structure in addition to modularity in the processing structure. Due to distributed control, a variety of operations can be performed without the need of glue logic for the interface of the chips