Using spin images for hand gesture recognition in 3D point clouds

Depth information from a captured scene can provide a more complete description of objects which can be exploited for recognition purposes. In this paper we propose a new approach to static hand gesture recognition that explores only depth information acquired with a Microsoft Kinect camera. Firstly, the regions of interest are extracted from depth data and a 3D point cloud is created taking into consideration the capture device settings. Then, for each 3D point in the simplified point cloud the local spin image descriptor is computed. The performance of hand pose detection greatly depends on the 3D shape descriptor used. KPCA is used for dimensionality reduction and for noise elimination from obtained spin image histograms. The most relevant features in terms of principal components are served as input for a SVM classifier. Experimental results show a high level of accuracy in recognizing static hand gestures using spin images in 3D point clouds.