A Feature Map Approach to Pose Estimation Based on Quaternions

This paper proposes a novel solution to the problem of pose estimation of three-dimensional objects using feature maps. Our approach relies on quaternions as the mathematical representation of object orientation. We introduce the rigid map, which is derived from Kohonen's self-organizing feature map. Its topology is fixed and chosen in accordance with the quaternion representation. The map is trained with computer-generated object views such that it responds to a preprocessed input image with one or more sets of object orientation parameters. Experimental results demonstrate that a pose estimate within the accuracy requirements can be found in more than 90% of all cases. Our current implementation operates at near frame rate on real input images.