Texture-less object detection and 6D pose estimation in RGB-D images

We present a cascade-like texture-less object detection and 6D pose estimation method exploiting both depth and color information from the RGB-D sensor. This is accomplished through both an offline and online phase. During the offline phase, a set of rendered templates is created from uniformly distributed sampling viewpoints obtained by employing the model of electrostatic charge distribution. During the online phase, our method converts sliding windows to scale-invariant RGB-D patches and employs a hash voting-based hypothesis generation scheme to compute a rough 6D pose hypothesis. Particle swarm optimization is employed to refine the 6D pose of the target object. We evaluate the algorithm against three datasets with the result that this method achieves high precision and good performance under various conditions. It was also applied to augmented reality and intelligent robotic manipulation yielding robust detection results.