Hand Orientation Regression Using Random Forest for Augmented Reality

We present a regression method for the estimation of hand orientation using an uncalibrated camera. For training the system, we use a depth camera to capture a large dataset of hand color images and orientation angles. Each color image is segmented producing a silhouette image from which contour distance features are extracted. The orientation angles are captured by robustly fitting a plane to the depth image of the hand, providing a surface normal encoding the hand orientation in 3D space. We then train multiple Random Forest regressors to learn the non-linear mapping from the space of silhouette images to orientation angles. For online testing of the system, we only require a standard 2D image to infer the 3D hand orientation. Experimental results show the approach is computationally efficient, does not require any camera calibration, and is robust to inter-person shape variation.