Pyramidal architecture for stereo vision and motion estimation in real-time fpga-based devices

Resumo

In this work, an exhaustive study for a low level vision engine is presented. Our hypothesis is based on recent development and advantages of FPGA devices (reduced power consumption, high processing capabilities) and the performance of new HDLs and synthesis tools. Thus we address with these powerful means the novel target of a low level vision engine on the same chip. The study aims to demonstrate that is possible the integration of multiple complex algorithms thanks to a proper adaptation and good design techniques. In particular we focus our architecture to a fine grain pipeline in opposition to the multi-core approach largely used in last architectures. Our approach benefits the power consumption and the size of the final implementation providing a very competitive system useful for industrial, robotic and research fields. For the first time is afforded a multi-scale and a multi-orientation optical flow and stereo on FPGA. The iterative nature of this approach degrades the processing speed but achieves an important accuracy and significantly enhances the working range. Final results in synthetic and real sequences demonstrate the competitive performance of the presented system.