Hardware-Software Codesign of Histogram of Oriented Gradients on Heterogeneous Computing Platform

Abstract

Histogram of oriented gradients (HOG) is a highly important feature representation in computer vision for many applications such as objection detection. The HOG computes local histograms of oriented gradients of pixel luminance on a dense grid of uniformly spaced cells and normalized to be a feature vector. Its computational complexity is high, and its implementation on edge computing and embedded devices is challenging. This paper proposes a hardware software codesign strategy to redesign the HOG algorithm. Pipelining and hardware acceleration by FPGA are applied in the design to the performance improvement of HOG. The design is implemented on a heterogeneous computing platform and with high level synthesis techniques exploiting C-code to accelerate the design of hardware circuits. Our results of full HD images achieve 500 times speed-up compared with software implementation.