Uncertainty of features in planar object recognition and a new classifier

Abstract

Stereoscopy has been widely used for passively estimating 3D information about a scene. There exist, however, many error sources such as sensor, quantization, segmentation, mismatching and so on, which induce errors in estimating the 3D information. These errors in turn cause errors in computing 3D features of objects to be recognized based on the 3D information. In this work, we analyze numerically the uncertainty of features of several 3D planar objects, and it is observed that the probability distribution of the features is similar to that of a gaussian random vector and varies with the position and orientation of objects. From these results we propose an adaptive algorithm to recognize 3D planar objects. Simulations show that, compared to the conventional minimum distance classifier, our proposed algorithm reduces the probability of misclassification. And we conclude that the uncertainty of the estimated 3D information should be taken into account for recognizing 3D objects more accurately.