Optimal Model-Based 6-D Object Pose Estimation With Structured-Light Depth Sensors

Structured-light (SL) depth sensors are widely used because of their simplicity in design and ability to process depth data with minimal computational expense. Certain SL light coding methods can, however, lead to a loss of information, as well as inhomogeneous depth errors that depend on the composition and properties of the scene. This results in a reduction of potential accuracy for model-based pose estimation methods that operate on the depth images or subsequently transformed three-dimensional point clouds, such as the popular class of point set registration (PSR) methods. We therefore formulate an asymptotically optimal maximum likelihood estimation (MLE) method that operates directly on the raw SL infrared (IR) images. The proposed SLIR-MLE method maximizes the likelihood of the measured IR image over the pose region given the object model, sensor model, and calibrated speckle and thermal noise distributions. We also formulate a method to compute the Fisher information contained in the IR image and resulting Cramér-Rao bound (CRB) of any unbiased pose estimator for unique SL sensor measurement data. SLIR-MLE is shown to nearly achieve the calculated CRB for the Kinect sensor by operating on the more informative raw IR images. Furthermore, our method is shown to outperform two cutting edge PSR methods by an order of magnitude in the respective mean square errors.