Tracking 3D shapes in noisy point clouds with Random Hypersurface Models

Depth sensors such as the Microsoft Kinect™ depth sensor provide three dimensional point clouds of an observed scene. In this paper, we employ Random Hypersurface Models (RHMs), which is a modeling technique for extended object tracking, to point cloud fusion in order to track a shape approximation of an underlying object. We present a novel variant of RHMs to model shapes in 3D space. Based on this novel model, we develop a specialized algorithm to track persons by approximating their shapes as cylinders. For evaluation, we utilize a Kinect network and simulations based on a stochastic sensor model.