A real-time computer vision library for heterogeneous processing environments

Abstract

With a variety of processing technologies available today, using a combination of different technologies often provides the best performance for a particular task. However, unifying multiple processors with different instruction sets can be a very ad hoc and difficult process. The Open Component Portability Infrastructure (OpenCPI) provides a platform that simplifies programming heterogeneous processing applications requiring a mix of processing technologies. These include central processing units (CPU), graphics processing units (GPU), field-programmable gate arrays (FPGA), general-purpose processors (GPP), digital signal processors (DSP), and high-speed switch fabrics. This thesis presents the design and implementation of a computer vision library in the OpenCPI framework, largely based on Open Source Computer Vision (OpenCV), a widely used library of optimized software components for real-time computer vision. The OpenCPI-OpenCV library consists of a collection of resource-constrained C language (RCC) workers, along with applications demonstrating how these workers can be combined to achieve the same functionality as various OpenCV library functions. Compared with applications relying solely on OpenCV, analogous OpenCPI applications can be constructed from many workers, often resulting in greater parallelization if run on multi-core platforms. Future OpenCPI computer vision applications will be able to utilize these existing RCC workers, and a subset of these workers can potentially be replaced with alternative implementations, e.g. on GPUs or FPGAs.