@article{oai:ipsj.ixsq.nii.ac.jp:00202719,
 author = {武山, 洪二郎 and 加藤, 武男 and 後藤, 邦博 and Kojiro, Takeyama and Takeo, Kato and Kunihiro, Goto},
 issue = {1},
 journal = {情報処理学会論文誌},
 month = {Jan},
 note = {本研究では高度運転支援アプリケーションへの適用を目的とし，LiDARより安価な単眼カメラと地図情報を用いて，誤差0.3m相当の高精度位置推定技術の実現を目指す．車載カメラ画像と地図による位置推定は，カメラ画像中の特徴と事前に生成した地図情報に含まれる特徴との照合を行うことで地図上の自車位置を推定する技術であるが，周囲に特徴物が少ない場合など，照合できる特徴の数が減少した場合に位置精度劣化の恐れがある．そこで本研究では，車載カメラ画像の時系列データを利用することで照合の手がかりを増加させる手法を提案し，環境変化に対する位置精度の頑健性を向上させた．実験では実走行データを用いて，照合できる特徴の数と位置精度との関係を検証した結果，提案手法では照合できる特徴数の減少にともなう位置精度劣化の度合いが大幅に改善される傾向が見られた．実環境に則した精度劣化シーン（構造物少，照明変化）では，位置誤差0.3m以下を満たす場所の割合は従来手法でそれぞれ70%台であったのに対し，提案手法では100%近くまで改善可能であることを確認した．, This study proposes a method to provide the localization accuracy within 0.3m for the land vehicles. The localization via feature point matching between monocular camera and the pre-built map has a potential to achieve desi-meter accuracy without using expensive sensors such as LiDAR. The accuracy of localization is stable when the number of matched feature points is sufficient, however in a scene where the feature point matching is difficult to be performed due to the illumination change, the accuracy of localization can be degraded since the decrease of matched feature points makes the localization accuracy unstable. The proposed method uses feature points in the sequential time-series images to pseudo-increase the number of feature point matched with the map, which improves the robustness against the environment where the feature point matching is difficult to perform. The experiment showed the result that the proposed method has improved the accuracy of the localization when the number of matched feature points is reduced. Evaluation in the real environment with the illumination change or lack of texture, the availability of localization within 0.3m accuracy has been improved to around 100% with the proposed method while that of the conventional method is around 70%.},
 pages = {16--25},
 title = {特徴点地図と単眼カメラ画像の時系列照合によるロバスト位置推定手法の提案},
 volume = {61},
 year = {2020}
}