講演情報
[6B-05]Incorporating Temporal Dynamics and Intricate Spatial-Temporal
Dependencies for ActionFormer
*Zhao Kunpeng1、大北 剛1、宮崎 朝陽1 (1. 九州工業大学知能情報工学研究系)
発表者区分：学生
論文種別：ロングペーパー
インタラクティブ発表：あり
論文種別：ロングペーパー
インタラクティブ発表：あり
キーワード：
Transformer、Human Activity Recognition、Deep Learning
Human Activity Recognition (HAR) relies on extracting meaningful patterns from sensor or visual data to classify activities. This paper introduces the CE-HAR framework, which incorporates channel-wise enhancement into Transformer-based architectures to improve HAR performance. Two modules, Adaptive Channel-wise Enhancement (AdapCE) and MaxPool Channel-wise Enhancement (MaxCE), are developed to address the challenges of high temporal dynamics and intricate spatial-temporal dependencies in inertial and visual features, respectively. Experiments on the WEAR dataset show that CE-HAR achieves substantial improvement, with a 16.01\% improvement in average mAP for inertial data and a 7.8\% improvement for visual data compared to the baseline ActionFormer. Additional evaluations on benchmark datasets confirm the robustness and effectiveness of the proposed approach. Our findings highlight the importance of channel-wise information processing and open new avenues for advancing HAR techniques.