A Feature Map Approach to Real-Time 3-D Object Pose Estimation from Single 2-D Perspective Views

A novel approach to the computation of an approximate estimate of spatial object pose from camera images is proposed. The method is based on a neural network that generates pose hypotheses in real time, which can be refined by registration or tracking systems. 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 orienta­ tion parameters. The key concepts proposed are representations of spatial orienta­ tion that result in continuous distance measures, and the choice of a fixed network topology that is best suited to the representation of 3-D orientation. Experimen­ tal results from both simulated and real images demonstrate that a pose estimate within the accuracy requirements can be found in more than 90% of all cases. The current implementation operates at near frame rate on real world images.