Presentation Information
[B-1C-07]A Learning Framework for Orientation-Invariant Human Action Recognition Using Multi-Vantage Micro-Doppler
〇YUPENG WANG1, Keita Nishi1, Minseok Kim1 (1. Niigata University)
Keywords:
Integrated Sensing and Communication,Human Activity Recognition,Multi-vantage,orientation-invarian,Machine learning
Integrated Sensing and Communication (ISAC) has emerged as a promising paradigm for next-generation wireless systems, where Human Activity Recognition (HAR) based on micro-Doppler signatures is one of its key sensing applications\cite{ISAC}. In practical HAR scenarios, however, the relative orientation between the human body and the sensing device is usually not fixed, causing significant variations in the observed micro-Doppler patterns. Multi-vantage observation can provide complementary motion information and improve robustness against such orientation-dependent effects, but deploying many antennas or conducting full multi-directional measurements is often impractical due to hardware complexity and implementation. To address this issue, this work proposes a learning framework for orientation-invariant HAR trained on multi-vantage micro-Doppler measurements. Based on real-world micro-Doppler measurements collected using a four-node multi-link measurement setup at two different subject-facing angles, we constructed an equivalent eight-link monostatic dataset and used it to train an orientation-invariant HAR model. We then evaluated the model under individual monostatic radar configurations.