Robust Model Based Tracking Using Edge Mapping And Refinement

We present a markerless tracking approach for augmented reality in poorly textured environments. The approach enables a robust and accurate camera pose estimation merely on basis of a coarse edge model. The edge model of the object to be tracked is enhanced and refined during the tracking process. New edges are added to the edge model and already existing ones are refined. A collection of reference poses with a set of corresponding edges, called keyposes, enables a selection of good edges to track depending on the current view and makes the tracking process robust and accurate. Keyposes are also used to reinitialize automatically after tracking failures, e.g. the object to be tracked is occluded. Therefore, the proposed method overcomes the limitations of traditionally used edge based tracking approaches in terms of reinitialization and edge model creation. Evaluation on synthetic and real image sequences demonstrates the significant improvement of the proposed method over a standard edge based tracking.