CSI-Former: Pay More Attention to Pose Estimation with WiFi.

Area of Science:

• Computer Vision
• Signal Processing
• Machine Learning

Background:

• Human pose estimation is crucial for various applications.
• Traditional image-based methods fail in poor lighting or darkness.
• Existing sensor-based methods (LiDAR, RF) are costly and require specialized equipment.

Purpose of the Study:

• To propose a novel, cost-effective WiFi-based human pose estimation method.
• To introduce the CSI-former architecture for WiFi-based pose estimation.
• To establish a new dataset (Wi-Pose) for evaluating WiFi-based pose estimation.

Main Methods:

• Utilized WiFi Channel State Information (CSI) for pose estimation.
• Developed a novel deep learning architecture, CSI-former, incorporating multi-head attention.
• Created and utilized the Wi-Pose dataset, comprising WiFi CSI, images, and skeleton annotations.

Main Results:

• CSI-former significantly enhances performance in wireless pose estimation.
• The proposed method outperforms traditional image-based pose estimation.
• The Wi-Pose dataset enables robust evaluation of WiFi-based pose estimation techniques.

Conclusions:

• CSI-former offers a promising solution for robust human pose estimation using WiFi signals.
• The developed Wi-Pose dataset facilitates future research in this domain.
• This work paves the way for accessible and effective wireless pose estimation.