Skip to Main Content
In model-based camera tracking where camera poses are estimated in such a way that projections of edges on a known 3D scene/object model are aligned with close and strong edges detected in camera images, a projection usually has multiple candidate correspondences (or hypotheses) and there is little information on which one is the true hypothesis. This ambiguity makes model-based camera tracking unstable and inaccurate. Therefore, this paper proposes an adaptive edge detection method that models the gradients of true hypotheses as a mixture of Gaussian distributions, adjusts the parameters of an edge detector based on the model, and selectively eliminates false hypotheses. In our preliminary experiments, the method reduced the pose error and jitter of a testbed model-based camera tracking system by 27% and 2%, respectively1.