3次元モデルからの物体認識特徴深索手順の自動生成

概要

論文の詳細を見る
We are developing a vision system for an intelligent robot which can efficieently recognize many kinds of object. This system consists of two parts ; automatic generation of an object recognition procedure from a 3-D model (off-line part), feature extraction and object recognition by executing the procedure (on-line part) . This paper presents the off-line part of the system.<BR>How to deal with the variations of object appearnces is the main issue in 3-D object recognition. In our approach, first, object appearances for various view directions are described with 2-D features which are derived from projections of the model. Then, the appearances for different view directions are analyzed by comparing descriptions for each direction, and a tree-like graph is constructed, where general descriptions visible in many directions are linked to more specific descriptions visible in fewer directions. Finally, the search order in the graph is determined by attaching a priority to each node of the graph. The capability of available image processing module in extracting features as well as the feature properties is considered in this priority evaluation<BR>The generated tree-like graph is used as an object recognition procedure. The system searches through the graph in the order of the priority attached to each node and looks for the features in the description from an image.
一般社団法人日本ロボット学会の論文