Hand pose estimation using HOG features from RGB-D data

Visual based recognition of hand gestures has been an active research field in recent years due to its efficiency in helping us achieve a more natural human-computer interaction. This paper presents a new approach to hand pose estimation using combined visual and geometric information obtained in a synchronized format from a RGB-D sensor. Firstly, we track the contour of the hand and recognize the fingertip positions. Then, Kernel Principal Component Analysis is used for selecting the most relevant elements from the histograms of oriented gradients feature vectors obtained on both color and depth data. We define an observation model based on the found fingertip positions and the selected principal components and we feed it as input for a Support Vector Machine classifier. Experimental results for the proposed method show that good detection percentages can be obtained with a small training dataset of real hand images and depth masks.