Flexible and efficient perspective transform engine

Abstract

Perspective transform (or Homography) is commonly used in image and computer vision processing, which typically solved using a back mapping algorithm. It is used in multiple use-cases e.g. view point change, fish-eye lens distortion correction, chromatic aberration correction, stereo image pair alignments etc. Prior work in literature addresses individual use-case by selecting data format, interpolation method and block sizes due to constraints of irregular data fetch. In this paper, we propose an area efficient and flexible solution for perspective transform engine to address multiple use-cases with software controlled data format, interpolation method and block size. The proposed solution consists of multiple innovative techniques namely On-The-Fly (OTF) transform, Off-line mesh based Look up Table (LUT) with optional down-sampling, corner pre-fetch and dual interpolation. The proposed solution achieves clock up- to 300 MHz with output pixel throughput of 1 cycle/pixel and 0.21 mm2 of silicon area in 28nm LP-CMOS process node.