TY - GEN

T1 - Hand gesture recognition using Histogram of Oriented Gradients and Partial Least Squares regression

AU - Misra, Arindam

AU - Takashi, Abe

AU - Okatani, Takayuki

AU - Deguchi, Koichiro

PY - 2011/12/1

Y1 - 2011/12/1

N2 - In this paper we propose a real-time hand gesture recognition system that employs the techniques developed for pedestrian detection to recognize a small vocabulary of human hand gestures. Our feature set comprises of grids of Histogram of Oriented Gradient (HOG) descriptors, with fine orientation binning and multi-level spatial binning for getting descriptors at the small as well as large scale. The overlapping descriptor blocks, which are contrast normalized to handle illumination changes, have a high degree of multicollinearity, resulting in a feature set of high dimensionality (more than 8000 dimensions), rendering it unsuitable for classification using the classical machine learning algorithms. Thus, we employ Partial Least Squares (PLS) regression as a 'class aware' method of dimensionality reduction, to project the feature vectors on to a lower dimensional space of 10 dimensions. We examine the results obtained by PLS as well as Principal Component Analysis (PCA) which show, that PLS outperforms PCA, and gives a better projection which preserves significant discriminative information.

AB - In this paper we propose a real-time hand gesture recognition system that employs the techniques developed for pedestrian detection to recognize a small vocabulary of human hand gestures. Our feature set comprises of grids of Histogram of Oriented Gradient (HOG) descriptors, with fine orientation binning and multi-level spatial binning for getting descriptors at the small as well as large scale. The overlapping descriptor blocks, which are contrast normalized to handle illumination changes, have a high degree of multicollinearity, resulting in a feature set of high dimensionality (more than 8000 dimensions), rendering it unsuitable for classification using the classical machine learning algorithms. Thus, we employ Partial Least Squares (PLS) regression as a 'class aware' method of dimensionality reduction, to project the feature vectors on to a lower dimensional space of 10 dimensions. We examine the results obtained by PLS as well as Principal Component Analysis (PCA) which show, that PLS outperforms PCA, and gives a better projection which preserves significant discriminative information.

UR - http://www.scopus.com/inward/record.url?scp=84872528352&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=84872528352&partnerID=8YFLogxK

M3 - Conference contribution

AN - SCOPUS:84872528352

SN - 9784901122115

T3 - Proceedings of the 12th IAPR Conference on Machine Vision Applications, MVA 2011

SP - 479

EP - 482

BT - Proceedings of the 12th IAPR Conference on Machine Vision Applications, MVA 2011

T2 - 12th IAPR Conference on Machine Vision Applications, MVA 2011

Y2 - 13 June 2011 through 15 June 2011

ER -