Using novel shape, color and texture descriptors for human hand detection

In this paper, we present a robust feature set to detect human hands in still images having simple as well as complex backgrounds. Our method relies on using a blend of existing and new shape-based, color-based and texture-based features. First, we identify the shortcomings of two existing features: Histograms of Oriented Gradient (HOG) and Color Name (CN). For HOG, we investigate the scenarios where the traditional block normalization schemes generate noisy results in near uniform regions in the image background and impede the accurate detection of human hands. We offer a more effective block normalization scheme for our new shape-based feature, αHOG, which results in considerably improved detection. Our new color-based feature, Clipped Color Name (CCN), caters for the noise induced color labels encountered in the CN feature, by modifying the probability assignment method for the basic colors in each pixel. For capturing the texture cues, we employ Local Binary Patterns (LBP) and Local Trinary Patterns (LTP). We compare the relative performance of the individual features in isolation and in different feature sets. For feature sets' comparison, the issue of high dimensional feature space generated as a result of feature fusion is addressed by using Partial Least Squares (PLS) for dimensionality reduction. Subsequently, we employ the non-linear Radial Basis Function Support Vector Machine (RBF SVM) classifier on PLS reduced feature sets. In our experiments, we use two different image datasets, namely the benchmark Cambridge Gesture Dataset (having simple backgrounds) and our own dataset (having a wider variety of complex backgrounds). Based on the experimental results, we find that out of the four feature sets we use, the feature set consisting of αHOG, CCN and LTP gives the best results in terms of the combined criteria of classification accuracy and computation time, and also offers improvement over the feature set proposed by Hussain and Triggs [- ].