画像の学習に基づく姿勢推定手法の学習条件による精度変化の検討

Okugawa Hiroyuki

本研究では,画像の学習に基づく姿勢推定手法の推定精度を検討する.検討の具体的な内容は,学習条件である学習画像枚数や固有空間の次元数を変化させ,その姿勢推定誤差を物体毎に比較することである.検討の結果,学習画像枚数が多い場合より少ない場合の方が精度が良くなる物体が存在した.そのような物体では,相関が高い画像同士が学習画像に含まれると推定精度が向上するという結果を得た.また,固有空間の次元は,物体により精度を保つのに必要な次元数が存在した.

In this research, the presumption accuracy of the posture presumption technique based on the study of the image is examined. A concrete content to examine is to change the number of sheets of the study image that is the study condition and the number of dimension of a peculiar space, and to compare the posture presumption error margins of each object. As a result of the examination, the object to which accuracy improved existed, when the number of sheets of the study image is a little in a lot of cases that are. In such an object, the result that the presumption accuracy improved when the image with a high correlation was included in the study image mutually was obtained. Moreover, the dimension of a peculiar space had a number of dimension necessary to keep accuracy with the object.

物体認識

EbC法

精度変化

Electrical engineering [ 540 ]

jpn

Copyright (c) 2008 Author

工学部学士

Engineering

広島大学