Real-time pose estimation of 3D objects from camera images using neural networks

This paper deals with the problem of obtaining a rough estimate of three dimensional object position and orientation from a single two dimensional camera image. Such an estimate is required by most 3-D to 2-D registration and tracking methods that can efficiently refine an initial value by numerical optimization to precisely recover 3-D pose. However the analytic computation of an initial pose guess requires the solution of an extremely complex correspondence problem that is due to the large number of topologically distinct aspects that arise when a three dimensional opaque object is imaged by a camera. Hence general analytic methods fail to achieve real-time performance and most tracking and registration systems are initialized interactively or by ad hoc heuristics. To overcome these limitations we present a novel method for approximate object pose estimation that is based on a neural net and that can easily be implemented in real-time. A modification of Kohonen´s self-organizing feature map is systematically trained with computer generated object views such that it responds to a preprocessed image with one or more sets of object orientation parameters. The key idea proposed here is to choose network topology in accordance with the representation of 3-D orientation. Experimental results from both simulated and real images demonstrate that a pose estimate within the accuracy requirements can be found in more than 81% of all cases. The current implementation operates at 10 Hz on real world images