Integer Arithmetic Approximation of the HoG Algorithm used for Pedestrian Detection

Srđan Sladojević1, Andraš Anderla1, Dubravko Ćulibrk1, Darko Stefanović11 and Bojan Lalić1

  1. Faculty of Technical Sciences, University of Novi Sad
    Trg D. Obradovića 6, 21000 Novi Sad, Serbia
    {sladojevic, andras, dculibrk, darkoste, blalic}@uns.ac.rs

Abstract

This paper presents the results of a study of the effects of integer (fixed-point) arithmetic implementation on classification accuracy of a popular open-source people detection system based on Histogram of Oriented Gradients. It is investigated how the system performance deviates from the reference algorithm performance as integer arithmetic is introduced with different bit-width in several critical parts of the system. In performed experiments, the effects of different bit-width integer arithmetic implementation for four key operations were separately considered: HoG descriptor magnitude calculation, HoG descriptor angle calculation, normalization and SVM classification. It is found that a 13-bit representation of variables is more than sufficient to accurately implement this system in integer arithmetic. The experiments in the paper are conducted for pedestrian detection and the methodology and the lessons learned from this study allow generalization of conclusions to a broader class of applications.

Key words

computer vision, fixed-point, histogram of oriented gradients, pedestrian detection

Digital Object Identifier (DOI)

https://doi.org/10.2298/CSIS160229011S

Publication information

Volume 14, Issue 2 (June 2017)
Year of Publication: 2017
ISSN: 1820-0214 (Print) 2406-1018 (Online)
Publisher: ComSIS Consortium

Full text

DownloadAvailable in PDF
Portable Document Format

How to cite

Sladojević, S., Anderla, A., Ćulibrk, D., Stefanović1, D., Lalić, B.: Integer Arithmetic Approximation of the HoG Algorithm used for Pedestrian Detection. Computer Science and Information Systems, Vol. 14, No. 2, 329-346. (2017)