2A1-J07 エッジパターン特徴とハイライト特徴を用いた机上透明物体認識(ロボットビジョン(1))

Abstract

This paper describes a method of recognition for transparent objects on the desk, using a monocular camera. First, we apply a machine leaning-based method using edge patterns to the transparent desktop object recognition. Edge patterns which appear in the local area of the image are represented by a probabilistic topic model. Edge pattern features indicating presences of specific edge patterns, are used for training a Support Vector Machine (SVM) binary classifier, which dscriminates between transparent objects and backgrounds. Then, we try to detect transparent objects in desk scenes using the classifier. Next, we define a highlight pattern feature describing highlight regions of objects. Highlight pattern features are represented by binarization of highlighted pixel ratio in local areas, and combined with edge pattern features to enhance the previous method. Finally, combined features are tested in cross-validation and object detection and shown to be superior to the previous edge pattern features.