| Real-time imaging applications, such as robotics, military target tracking, autonomous vehicle control, and on-line video processing, require huge computational performance due to the large datasets involved. For this reason, real-time imaging systems have traditionally been built from custom hardware designed to perform specific algorithms. However, custom hardware implementations are expensive, and the real-time performance is achieved at the cost of end-user programmability and flexibility. A more generalized flexible approach is to parallelize the computations by taking advantage of their inherent concurrency and to utilize a scalable parallel hardware architecture.;It is clear that developing a general real-time parallel image processing system involves much more than building a high performance parallel hardware architecture. Without high-level programming environments and tools designed specifically for building parallel imaging software, parallel architectures are difficult to program, and thus not cost-effective solutions.;The goal of this dissertation has been to create a flexible and easy to use image processing programming environment for generating real-time parallel software implementations. I show that, through the use of Model-Integrated Program Synthesis (MIPS), parallel implementations of image processing data flows can be synthesized from high-level graphical specifications. The complex details inherent to parallel software development become transparent, enabling the cost-effective exploitation of parallel hardware for building more flexible and powerful real-time imaging systems.;As proof of concept, I present MIRTIS (Model Integrated Real-Time Image Processing System). MIRTIS employs the Multigraph Architecture (MGA), a framework and set of tools for building MIPS systems, to generate parallel run-time kernel on a network of Texas Instruments TMS320C40 Digital Signal Processors. The MIRTIS system provides a graphical model builder for declaring (1) the computations to be performed, (2) the available hardware resources, and (3) the performance constraints of the application. A model interpreter automatically decomposes, scales, and allocates the computations to the available resources, then generates a real-time parallel C40 implementation.;MIRTIS is a clear example of how the MGA can be used to synthesize parallel real-time image processing systems which are cost-effective to program, configure, and scale. |
