Preliminary Investigation of Packing Process Recognition Using Cross Knowledge Distillation
アイテムタイプ
SIG Technical Reports
言語
英語
キーワード
機械学習・知識蒸留手法
著者所属
Graduate School of Information Science and Technology, Osaka University
Graduate School of Information Science and Technology, Osaka University
Graduate School of Information Science and Technology, Osaka University
著者所属(英)
Graduate School of Information Science and Technology, Osaka University
Graduate School of Information Science and Technology, Osaka University
Graduate School of Information Science and Technology, Osaka University
Human activity recognition (HAR) often utilizes inertial measurement unit (IMU) data due to its effectiveness in various environments. However, IMU data lacks spatial position information and environmental context, limiting its robustness against changes in motion patterns, such as speed variations. Skeleton data, while providing detailed spatial information and better capturing human motion dynamics, can lead to serious occlusion problems in industrial environments with complex surroundings. To address these challenges, we propose a novel cross-knowledge distillation method, a machine-learning technique that transfers knowledge from a larger teacher model to a smaller student model. This approach enhances a student model based on IMU data through cross-modal knowledge distillation from a skeleton-based model. By leveraging the skeleton-based model to extract insights on capturing spatial and temporal information on the body in time-series data and superior prediction capabilities, the student model is significantly improved in its ability to capture intricate and rapidly evolving motion patterns. This method enables the student model to assimilate salient features from the teacher model, thereby enhancing overall performance in HAR tasks.
論文抄録(英)
Human activity recognition (HAR) often utilizes inertial measurement unit (IMU) data due to its effectiveness in various environments. However, IMU data lacks spatial position information and environmental context, limiting its robustness against changes in motion patterns, such as speed variations. Skeleton data, while providing detailed spatial information and better capturing human motion dynamics, can lead to serious occlusion problems in industrial environments with complex surroundings. To address these challenges, we propose a novel cross-knowledge distillation method, a machine-learning technique that transfers knowledge from a larger teacher model to a smaller student model. This approach enhances a student model based on IMU data through cross-modal knowledge distillation from a skeleton-based model. By leveraging the skeleton-based model to extract insights on capturing spatial and temporal information on the body in time-series data and superior prediction capabilities, the student model is significantly improved in its ability to capture intricate and rapidly evolving motion patterns. This method enables the student model to assimilate salient features from the teacher model, thereby enhancing overall performance in HAR tasks.
書誌レコードID
AA12628327
雑誌名
研究報告コンシューマ・デバイス＆システム（CDS）
巻
2024-CDS-41
号
3
ページ
1 - 8
発行年
2024-09-19
ISSN
2188-8604
Notice
SIG Technical Reports are nonrefereed and hence may later appear in any journals, conferences, symposia, etc.
Preliminary Investigation of Packing Process Recognition Using Cross Knowledge Distillation
IPSJ-CDS24041003.pdf
(1.71MB)
