Exposing the CSI (dataset)

This repository contains the links to the dataset + ground-truth videos for our paper Exposing the CSI: A systematic Investigation of CSI-based Wi-Fi Sensing Capabilities and Limitations.

Main features

The dataset contains CSI sequences collected in indoor environments while a target person performs different activities. The main features of the dataset are the following:

• 12 target activities, 80 seconds for each activity
• 7 different scenarios (3 people, 3 environments)
• CSI collected using 160-MHz 802.11ax devices with 4 antennas
• 3 Wi-Fi receivers collecting the same frames in different locations
• anonymized video ground-truth with reference keypoints

---

Dataset

We split our dataset into multiple repositories hosted on Zenodo. Each repository has size of approx 11 GB and contains the CSI data of a different scenario, from S1 to S7.

Scenario	Candidate	Environment	Day
S1	A	Lab	1
S2	B	Lab	1
S3	C	Lab	1
S4	A	Lab	1
S5	A	Lab	2
S6	A	Office	2
S7	A	Hall	2

Activities

This is the list of activities considered in every scenario:

A. Walk	D. Sitting	G. Wave hands	J. Wiping
B. Run	E. Empty room	H. Clapping	K. Squat
C. Jump	F. Standing	I. Lay down	L. Stretching

---

Videos

We recorded every experiment with a camera and used VideoPose3D to extract the reference keypoints of the people moving. The keypoints files are in the video_keypoints directory in this repository. They do not contain personal information about the candidates, but can be used to re-create an anonymized video ground truth of the experiments showing how the candidates are moving.

• Note #1: Keypoints for static activities (like sitting or laying down) are not included in this repository because there is no movement to show.

• Note #2: Unfortunately, videos for activities A and B for the scenario S3 are missing due to a data pre-processing issue.

Instructions to generate the videos

0. (optional) The system was tested on Ubuntu using Python 3.8 only, so it is suggested to install all the dependencies in a conda environment to avoid polluting your workspace.

conda create --name csi python=3.8
conda activate csi
pip install numpy torch matplotlib

1. Clone the VideoPose3D repository:

git clone https://github.com/facebookresearch/VideoPose3D

2. Download the pre-trained model in the checkpoint directory of VideoPose3D:

cd VideoPose3D
mkdir checkpoint
cd checkpoint
wget https://dl.fbaipublicfiles.com/video-pose-3d/pretrained_h36m_detectron_coco.bin
cd ..

3. Then, copy our keypoint files into the data directory of VideoPose3D:

cp /path/to/exposing-the-csi/video_keypoints/* data/

4. Finally, generate the anonymized video starting from the keypoints. Note: the following command generates the video for the activity A (walk) in the scenario S1 (see above). Change the value of the variable activity to generate the other videos accordingly.

activity=S1_A; python run.py -d custom -k ${activity} -arc 3,3,3,3,3 -c checkpoint --evaluate pretrained_h36m_detectron_coco.bin --render --viz-subject ${activity}.mp4 --viz-action custom --viz-camera 0 --viz-video data/blank.mp4 --viz-output ${activity}.mp4 --viz-size 6

---

References

If you find these resources useful, please consider citing our paper:

M. Cominelli, F. Gringoli and F. Restuccia, "Exposing the CSI: A Systematic Investigation of CSI-based Wi-Fi Sensing Capabilities and Limitations," 2023 IEEE International Conference on Pervasive Computing and Communications (PerCom), Atlanta, GA, USA, 2023, pp. 81-90, doi: 10.1109/PERCOM56429.2023.10099368.