Pose-insensitive nose detection in TOF-scans

One of the fundamental benefits of 3D range data, as opposed to 2D images, is the pose invariance of 3D geometrical information. Time-of-flight (TOF) cameras provide such range data at a high temporal, but so far only at a medium spatial resolution. We present an algorithm to detect the tip of the nose in TOF range data, where the subjects’ head pose varies between −90° and 90° around the vertical axis (yaw) and between −30° and 30° in pitch rotation. The main idea is to rotate the head around a virtual head axis and extract the silhouettes for all rotations in 5° steps. To find the correct rotation and the tip of the nose, we compare these extracted silhouettes to an average head profile template, which was calculated from 200 high resolution laser-scans. In a test set of 120 TOF scans of 10 different persons, the detected nose tip was within 10 mm from the correct position in 89% and within 25 mm in 97.5% of the cases.