Pointing hand distinction by improved HOG and wavelet multi-scale transform

Abstract

In order to improve the accuracy of distinguish left hands and right ones from a single fixed camera, this paper proposes a new method to deal with it. The method is based on the combination of improved histograms of oriented gradient (HOG) and wavelet multi-scale transform (WMT). By using multi-scale transform to extract the image edge information, the image global edge can be described well. HOG descriptor is improved to simplify the calculating procedure. We use it to extract the detail features of the image edge. Supported vector machine is used to classify left pointing hand, right pointing hand and negative one. Experiment result demonstrates that the proposal increases 23.1% in accuracy rate comparing with the performance of edge orientation histogram. Meanwhile, it is not only less time-consuming than that of HOG, but also has higher performance. © 2012 Springer-Verlag Berlin Heidelberg.