Customizing CPU Instructions for Embedded Vision Systems

Abstract

This paper presents the customization of two processors: the Altera NIOS2 and the Tensilica Xtensa, for fundamental algorithms in embedded vision systems: the salient point extraction and the optical flow computation. Both can be used for image stabilization, for drones and autonomous robots. Using 16-bit floating-point instructions, the architecture optimization is done in terms of accuracy, speed and power consumption. A comparison with a PowerPC Altivec is also done.