Three-dimensional object recognition is one of very important and interesting research topics in computer vision, and accurate 3D position information about objects in a scene to be recognized is indispensable for recognizing them. Unfortunately, it is well-known that it is impossible to find the true 3D information due to several error sources such as sensor, quantization, preprocessing (edge detection, line fitting, feature point extraction), mismatching and so on. In order to estimate the 3D information more accurately, a lot of algorithms and systems have been proposed so far.
And, it is evident that the features of objects obtained from this erroneous 3D position information become inaccurate, and that the range estimation error plays an important role on 3D object recognition.
In this dissertation, we analyze numerically the uncertainty of features of 3D planar objects. It is observed that the probability distribution of the numerical features is similar to that of gaussian random vector and varies with the pose (position and orientation) of objects. From these results we propose an adaptive aligorithm to recognize 3D polyhedral objects. Simulations show that, compared to the conventional minimum distance classifier, our proposed algorithm reduces the probability of misclassification. And we conclude that the uncertainly of the estimated 3D information should be taken into account for recognizing 3D polyhedral objects more accurately.