Robust Fitting of 3D CAD Models to Video Streams

We present a robust and accurate semi-automatic algorithm for registering and tracking a 3D geometric model in a 3D video stream. The algorithm is a generalization of the “Iterative Closest Point” technique. Each iteration is composed of two steps: computation of camera, parameters, and 3D/2D vertex matching. This last step is performed by polygon fitting in an edge image. To account for false matches, we use a robust M-estimation both for camera parameter estimation and 2D feature extraction. Experimental results show that accurate registration can be obtained even with very noisy outdoor images and incomplete data. Error analysis proves that. the accuracy is obtained at the pixel level.