Temporal Action Segmentation of Ultra-Low Frame Rate Excavator Work Video at the Construction Site Using Transformer-Based Model (ASFormer)

カテゴリ：部門大会

論文No：SS1-7

グループ名：【C】2025年電気学会電子・情報・システム部門大会

発行日：2025/8/20

タイトル(英語)：Temporal Action Segmentation of Ultra-Low Frame Rate Excavator Work Video at the Construction Site Using Transformer-Based Model (ASFormer)

著者名：Sereepookkana Kanok（）,Orachon Teerapong（KOSEN-KMITL, King Mongkut's Institute of Technology Ladkrabang, Thailand）,Doi Shigeo（苫小牧工業高等専門学校）,Itayama Yuichiro（株式会社 EARTHBRAIN）

著者名(英語)： Kanok Sereepookkana (),Teerapong Orachon (King Mongkut's Institute of Technology Ladkrabang),Shigeo Doi (National Institute of Technology, Tomakomai College),Yuichiro Itayama (EARTHBRAIN)

キーワード：Temporal Action Segmentation,Ultra-Low Frame Rate Video,Excavator Work Video,Egocentric Video,Transformer-Based ModelASFormer

要約(日本語)：Construction sites leverage technology to track on-site situations, enabling more efficient planning of construction projects. One popular method for tracking is by installing a camera on excavator machines to record construction work videos, as this approach is resource-efficient. These raw videos, referred to as untrimmed videos, are valuable because they can be analyzed to gain insights into on-site activities. To automatically analyze and convert untrimmed videos into segmented videos, the Temporal Action Segmentation (TAS) technique is employed. Our work focuses on a successive frame-level classification Transformer-based TAS model, ASFormer. Experimental results show that ASFormer can also be effectively applied to our excavator dataset, achieving a frame-wise accuracy of 83.13%.

本誌掲載ページ：1777-1779p

原稿種別：英語

PDFファイルサイズ：301Kバイト