2026-03-14T13:47:30Z https://ipsj.ixsq.nii.ac.jp/oai

oai:ipsj.ixsq.nii.ac.jp:00208653 2025-01-19T18:46:27Z 1164:4061:10116:10421

Preliminary investigation on activity recognition for packaging tasks using motif-guided attention networks Preliminary investigation on activity recognition for packaging tasks using motif-guided attention networks Jaime, Morales Naoya, Yoshimura Qingxin, Xia Takuya, Maekawa Atsushi, Wada Yasuo, Namioka Jaime, Morales Naoya, Yoshimura Qingxin, Xia Takuya, Maekawa Atsushi, Wada Yasuo, Namioka 行動識別 This study presents a method for recognizing packaging tasks using wrist-worn accelerometer sensors under real conditions. As the lead times and actions of packaging activities depend on the number of objects to pack along with the size and shape of each object, it is difficult to recognize operations during every period. We propose a segmentation neural network augmented with a multi-head attention mechanism to capture actions found in a specific operation, which can be useful to identify individual operations. To efficiently detect useful actions with limited training data, we propose an attention guiding approach based on existing motif detection algorithms, which find actions (motifs) that frequently appear in a specific operation. We then use the occurrence of these motifs as a target for each attention head, enabling it to increase its ability to recognize similar operations during the packaging process. We evaluate our framework using data obtained in an actual logistics center. This study presents a method for recognizing packaging tasks using wrist-worn accelerometer sensors under real conditions. As the lead times and actions of packaging activities depend on the number of objects to pack along with the size and shape of each object, it is difficult to recognize operations during every period. We propose a segmentation neural network augmented with a multi-head attention mechanism to capture actions found in a specific operation, which can be useful to identify individual operations. To efficiently detect useful actions with limited training data, we propose an attention guiding approach based on existing motif detection algorithms, which find actions (motifs) that frequently appear in a specific operation. We then use the occurrence of these motifs as a target for each attention head, enabling it to increase its ability to recognize similar operations during the packaging process. We evaluate our framework using data obtained in an actual logistics center. technical report 情報処理学会 2020-12-01 application/pdf 研究報告ユビキタスコンピューティングシステム（UBI） 11 2020-UBI-68 1 8 2188-8698 AA11838947 https://ipsj.ixsq.nii.ac.jp/record/208653/files/IPSJ-UBI20068011.pdf eng