{"links":{},"id":2008670,"metadata":{"_oai":{"id":"oai:ipsj.ixsq.nii.ac.jp:02008670","sets":["1164:4619:1767693301673:1772440944186"]},"path":["1772440944186"],"owner":"80578","recid":"2008670","title":["物理シミュレーションに基づく大規模事前学習による物理的妥当性を備えた歩行者軌跡予測"],"pubdate":{"attribute_name":"PubDate","attribute_value":"2026-03-17"},"_buckets":{"deposit":"6e139de4-a17e-42b2-83df-fdcacdb07337"},"_deposit":{"id":"2008670","pid":{"type":"depid","value":"2008670","revision_id":0},"owners":[80578],"status":"published","created_by":80578},"item_title":"物理シミュレーションに基づく大規模事前学習による物理的妥当性を備えた歩行者軌跡予測","author_link":[],"item_titles":{"attribute_name":"タイトル","attribute_value_mlt":[{"subitem_title":"物理シミュレーションに基づく大規模事前学習による物理的妥当性を備えた歩行者軌跡予測","subitem_title_language":"ja"}]},"item_keyword":{"attribute_name":"キーワード","attribute_value_mlt":[{"subitem_subject":"CVIM","subitem_subject_scheme":"Other"}]},"item_type_id":"4","publish_date":"2026-03-17","item_4_text_3":{"attribute_name":"著者所属","attribute_value_mlt":[{"subitem_text_value":"豊田工業大学"},{"subitem_text_value":"京都工芸繊維大学"},{"subitem_text_value":"豊田工業大学"}]},"item_4_text_4":{"attribute_name":"著者所属(英)","attribute_value_mlt":[{"subitem_text_value":"Toyota Technological Institute","subitem_text_language":"en"},{"subitem_text_value":"Kyoto Institute of Technology","subitem_text_language":"en"},{"subitem_text_value":"Toyota Technological Institute","subitem_text_language":"en"}]},"item_language":{"attribute_name":"言語","attribute_value_mlt":[{"subitem_language":"jpn"}]},"item_publisher":{"attribute_name":"出版者","attribute_value_mlt":[{"subitem_publisher":"情報処理学会","subitem_publisher_language":"ja"}]},"publish_status":"0","weko_shared_id":-1,"item_file_price":{"attribute_name":"Billing file","attribute_type":"file","attribute_value_mlt":[{"url":{"url":"https://ipsj.ixsq.nii.ac.jp/record/2008670/files/IPSJ-CVIM26245004.pdf","label":"IPSJ-CVIM26245004.pdf"},"date":[{"dateType":"Available","dateValue":"2028-03-17"}],"format":"application/pdf","billing":["billing_file"],"filename":"IPSJ-CVIM26245004.pdf","filesize":[{"value":"1.7 MB"}],"mimetype":"application/pdf","priceinfo":[{"tax":["include_tax"],"price":"660","billingrole":"5"},{"tax":["include_tax"],"price":"330","billingrole":"6"},{"tax":["include_tax"],"price":"0","billingrole":"20"},{"tax":["include_tax"],"price":"0","billingrole":"44"}],"accessrole":"open_date","version_id":"c63276e0-51ea-4ec7-9990-b0f0a4fbc0d8","displaytype":"detail","licensetype":"license_note","license_note":"Copyright (c) 2026 by the Information Processing Society of Japan"}]},"item_4_creator_5":{"attribute_name":"著者名","attribute_type":"creator","attribute_value_mlt":[{"creatorNames":[{"creatorName":"武次,広夢"}]},{"creatorNames":[{"creatorName":"延原,章平"}]},{"creatorNames":[{"creatorName":"浮田,宗伯"}]}]},"item_4_source_id_9":{"attribute_name":"書誌レコードID","attribute_value_mlt":[{"subitem_source_identifier":"AA11131797","subitem_source_identifier_type":"NCID"}]},"item_4_textarea_12":{"attribute_name":"Notice","attribute_value_mlt":[{"subitem_textarea_value":"SIG Technical Reports are nonrefereed and hence may later appear in any journals, conferences, symposia, etc."}]},"item_resource_type":{"attribute_name":"資源タイプ","attribute_value_mlt":[{"resourceuri":"http://purl.org/coar/resource_type/c_18gh","resourcetype":"technical report"}]},"item_4_source_id_11":{"attribute_name":"ISSN","attribute_value_mlt":[{"subitem_source_identifier":"2188-8701","subitem_source_identifier_type":"ISSN"}]},"item_4_description_7":{"attribute_name":"論文抄録","attribute_value_mlt":[{"subitem_description":"本研究では，歩行者軌跡予測における精度および物理的妥当性の向上を目的とした，物理シミュレーションに基づく事前学習手法を提案する．従来手法は観測長が十分な場合には高い精度を発揮するが，観測長が限られた状況では物理的に不自然な予測を出力することがある．このような予測の失敗は，歩行者が障害物の陰から突然現れる状況など，安全上重要な場面において危険性が高い．これまで，3次元人体姿勢を用いることで物理制約を暗黙的に学習する試みがあったが，高品質な3次元姿勢アノテーションを付与したデータには限りがあり，依然として物理的妥当性が欠如していた．そこで本研究では，物理シミュレーションを用いた大規模事前学習により，姿勢と軌跡の物理的関係性を学習する．具体的には，ヒューマノイドが目標軌跡に追従する歩行動作を大規模に物理シミュレーションし，その過程で得られる姿勢と軌跡のペアを用いて，軌跡予測とその物理的妥当性の評価を同時学習する．事前学習後は，軽量なアダプタのみを追加して転移学習することで，物理的知識を保持しつつデータセット固有のパターンや社会的相互作用を学習する．さらに，提案するTarget-relative Positional Encoding (TaPE)により，物理シミュレータのシングルエージェント環境で学習した表現をマルチエージェントの軌跡予測へ効果的に転移することを可能とした．実験では，JTA，JRDB，Urbanの3つのデータセットにおいて，提案手法による予測誤差の低減を確認した．特に，観測長が不十分で姿勢の理解がより重要となる短観測シナリオにおいて，従来手法と比較して平均予測誤差を最大50%低減することに成功した．","subitem_description_type":"Other"}]},"item_4_biblio_info_10":{"attribute_name":"書誌情報","attribute_value_mlt":[{"bibliographicPageEnd":"8","bibliographic_titles":[{"bibliographic_title":"研究報告コンピュータビジョンとイメージメディア（CVIM）"}],"bibliographicPageStart":"1","bibliographicIssueDates":{"bibliographicIssueDate":"2026-03-17","bibliographicIssueDateType":"Issued"},"bibliographicIssueNumber":"4","bibliographicVolumeNumber":"2026-CVIM-245"}]},"relation_version_is_last":true,"weko_creator_id":"80578"},"created":"2026-03-11T07:08:22.809694+00:00","updated":"2026-03-11T07:08:27.787949+00:00"}