REAL-CSI: Benchmarking DL Architectures for Wi-Fi CSI Human Activity Recognition


Emiroglu E. B., Çelik Y. C., ALİOĞLU A., ULUSOY İ.

34th Signal Processing and Communications Applications Conference, SIU 2026, İstanbul, Türkiye, 7 - 10 Temmuz 2026, (Tam Metin Bildiri)

  • Yayın Türü: Bildiri / Tam Metin Bildiri
  • Doi Numarası: 10.1109/siu71813.2026.11636516
  • Basıldığı Şehir: İstanbul
  • Basıldığı Ülke: Türkiye
  • Anahtar Kelimeler: benchmark, dataset, deep learning, human activity recognition, Wi-Fi CSI, wireless sensing
  • Orta Doğu Teknik Üniversitesi Adresli: Evet

Özet

This paper presents the REAL-LAB CSI HAR Dataset, a publicly available Wi-Fi Channel State Information (CSI) dataset for human activity recognition (HAR), together with a controlled deep learning benchmark. We contribute to two key gaps in prior work: the lack of synchronized, balanced datasets and the limited understanding of how architecture and model capacity influence performance. Using a microsecond-synchronized acquisition pipeline, we collect a multi-class dataset and evaluate seven neural architectures across four parameter scales under a multi-seed protocol. The results show that architectural inductive bias dominates model capacity: 2D convolutional networks achieve the best accuracy-efficiency trade-off (96.40%), while larger models often yield diminishing or negative returns. These findings highlight that effective CSI-HAR design depends more on architecture than scale. The dataset and benchmark provide a reproducible foundation for future research.