Novel Census Transform Hardware IP

• Main Authors: Reghunath, Anish, Mody, Mihir, Sanghvi, Hetul, Ankur
• Format: Conference Proceeding
• Language: English
• Subjects: Automotive; Census Transform; Computer vision; Hardware; Hardware IP; Heuristic algorithms; Industrial; Jacinto; Logic gates; Optical Flow; Robustness; Transforms; Very large scale integration
• ISSN: 2380-6923
• DOI: 10.1109/VLSID51830.2021.00007

Census Transform is a commonly used pixel transform technique in Computer Vision (CV) algorithms to provide robustness against illumination variance. Implementing this as a low gate-count and power-efficient HW IP is critical to support CV algorithms in SoCs targeting Automotive and Industrial markets. This paper gives an introduction to designing a HW IP for Census Transform as defined in the literature and proposes an efficient implementation by removing redundant computation by up-to 50% compared to the original algorithm. The paper also proposes a novel Census Transform algorithm by introducing a new pre-processing step which eliminates a key equality check during comparisons in the above implementation. The proposed novel algorithm results in a 27% lower gate-count and 13% lower power IP compared to the original implementation. The quality simulation of the proposed novel algorithm in the context of Optical Flow (an example CV algorithm) application shows only 1.48% flow vectors have differences compared to the original algorithm, resulting in a negligible impact to optical flow application. The proposed novel solution is implemented in Texas Instruments (TI) Jacinto7 series of product line.