High-speed Harris Corner Detection using residue number system

Abstract

Harris Corner Detector (HCD) is one of the common algorithms used in computer vision to locate the corner of an image for defining features. Corners are an important feature in computer vision because they are invariant to illumination change, translation or rotation, and image noise. This feature is widely used in motion detection, object tracking and recognition and stitching application. There are four main blocks in the Harris system, mask generation in spatial derivatives block, Gaussian filter block to remove unwanted noises, Harris response computation, and follow by non-maximum suppression block, which is used to compare and determine the corner pixel. HCD determine the corner by calculating the eigenvalue or change in intensity of shifting scanning window using Taylor series expansion. The Gaussian filtering is computationally expensive and requires a costly multiplier–accumulator (MAC unit). Therefore, the speed of HCD will be highly affected or limited by the MAC operation. The main objective of this project is to design a high speed HCD by implementing Residue number system (RNS) in the Gaussian filter block. The performance of the RNS-based HCD will be assess, benchmark with the MATLAB implementation of HCD and analyse using PPA (Power Performance Area). All the designs are synthesized using 180nm SilTerra library with fast database in Synopsys Design Compiler. In a conventional binary system, a large number will have a wider bit width and will require a larger arithmetic unit that slows down the computation. RNS is a fast arithmetic algorithm that is carry-free and can support high-speed and parallel arithmetic operations. RNS converts the larger number into a group of smaller numbers and performs the same math operation using a smaller math unit to speed up the design. The conventional binary number representation in the MAC unit of the computationally intensive Gaussian blocks was replaced by an RNS-based MAC unit. RNS with three moduli set, {2n-1,2n,2n+1}, n=9,13 provide the appropriate dynamic range to cover all the possible numbers in the kernel multiplication to prevent overflow. The CPD of the RNS-based Gaussian filter is reduced by 38%, with almost the same power consumption and an additional 21% area or gate count. The maximum operating frequency was increased by 60%, hence providing higher throughput than the conventional binary Gaussian filter. The results show that RNS is a good approach to improve computationally intensive applications such as digital filters and cryptographic applications with some additional areas.