Dynamically reconfigurable architecture for embedded computer vision systems

Abstract

The objective of this research work is to design, develop and implement a new architecture which integrates on the same chip all the processing levels of a complete Computer Vision system, so that the execution is efficient without compromising the power consumption while keeping a reduced cost. For this purpose, an analysis and classification of different mathematical operations and algorithms commonly used in Computer Vision are carried out, as well as a in-depth review of the image processing capabilities of current-generation hardware devices. This permits to determine the requirements and the key aspects for an efficient architecture. A representative set of algorithms is employed as benchmark to evaluate the proposed architecture, which is implemented on an FPGA-based system-on-chip. Finally, the prototype is compared to other related approaches in order to determine its advantages and weaknesses.