FallDeWideo: Vision-Aided Wireless Sensing Dataset for Fall Detection with Commodity Wi-Fi Devices

Abstract

Falling is one of the most dangerous events for the elderly and one of the most pressing public concern, which calls for an accurate, efficient, ubiquitous, cost-effective and privacy-preserving fall detection system to mitigate the negative consequences. Wi-Fi sensing-based method is considered as a potential technique to meet such needs. Existing solutions simply treat fall detection simply as a binary classification task, which leads to interpretability risks. To address this issue, we present FallDeWideo, the first multi-modal dataset dedicated to fall detection, comprising Wi-Fi CSI data and videos recorded during various kinds of events. We provide benchmark model for this dataset as well. Specifically, we train a CSI-based human pose estimation model (HPE) using video data as the supervision modality. The trained model can estimate human pose solely on Wi-Fi channel state information (CSI), and then detects whether a fall event occurs. This pipeline extracts rich information from CSI data and detects fall with more than just an alarm, but attached with an HPE, which allows more refined management of fall risks. We envision that this dataset will contribute to the wireless sensing research coomunity with respect to healthcare, action recognition, and cross-modal sensing. Codes and link to the dataset is available at https://github.com/shawnnn3di/falldewideo.