Datasets:

RS2002
/

WiCount

language:
  - en
task_categories:
  - time-series-forecasting
  - time-series-classification
tags:
  - wireless-sensing
  - csi
  - people-counting
  - wifi

WiCount

Dataset Description

Repository (WiCount subdirectory): CSI-BERT2/WiCount at main · RS2002/CSI-BERT2
Code: https://github.com/RS2002/CSI-BERT2
Paper: CSI-BERT2: A BERT-inspired Framework for Efficient CSI Prediction and Classification in Wireless Communication and Sensing, IEEE Transactions on Mobile Computing (TMC), 2025
Contact: zzhaock@connect.ust.hk
Collectors: Zijian Zhao, Tingwei Chen
Organization: AI-RAN Lab (hosted by Prof. Guangxu Zhu) in SRIBD, CUHK(SZ)
Dataset Summary: The WiCount dataset contains synchronized Channel State Information (CSI), Received Signal Strength Indicator (RSSI), and timestamp data collected using ESP32-S3 devices for WiFi-based people number estimation in a meeting room scenario. The dataset includes samples for estimating the number of people (0–3) in the environment.
Tasks: People Number Estimation

Sample Usage

To use this dataset with the CSI-BERT2 code, first clone the repository:

git clone https://github.com/RS2002/CSI-BERT2
cd CSI-BERT2

Then you can use the provided scripts for pre-training, fine-tuning, and inference. Replace <data path> with the path to the WiCount dataset downloaded from Hugging Face.

Pre-training

python pretrain.py --GAN --data_path <data path>

If you do not want to use the discriminator, you can omit the --GAN flag.

Fine-tuning for CSI Prediction

python prediction.py --GAN --data_path <data path> --parameters <fold path of the whole pre-trained models>

Fine-tuning for CSI Sensing Task (e.g., People Number Estimation)

For the WiCount dataset, use task "people".

python finetune.py --data_path <data path> --class_num <class num> --task "people" --path <parameter path of the backbone> --mode <mode>

The mode parameter can be set as 0, 1, or 2, corresponding to three experiments in the paper:

0: Training Set (100Hz), Testing Set (100Hz)
1: Training Set (100Hz+50Hz), Testing Set (100Hz+50Hz)
2: Training Set (100Hz), Testing Set (50Hz)

Inference for CSI Prediction

python prediction.py --data_path <data path> --parameters <fold path of the whole pretrained models> --eval_percent <the percentage of CSI sequence to be predicted>

Dataset Structure

Data Instances

Each instance is a .csv file representing a 60-second sample with the following columns:

seq: Row number of the entry.
timestamp: UTC+8 time of data collection.
local_timestamp: ESP32 local time.
rssi: Received Signal Strength Indicator.
data: CSI data with 104 numbers representing 52 subcarriers, where each subcarrier's complex CSI value is computed as a[2i] + a[2i+1]j.
Other columns: Additional ESP32 device information (e.g., MAC, MCS details).

Data Fields

Field Name	Description
seq	Row number of the entry
timestamp	UTC+8 time of data collection
local_timestamp	ESP32 local time
rssi	Received Signal Strength Indicator
data	CSI data (104 numbers, representing 52 subcarriers as complex values)
Other columns	Additional ESP32 metadata (e.g., MAC address, MCS details)

Data Splits

The dataset is organized by the number of people (0–3), with each folder containing .csv files corresponding to the number of people present in the environment:

Folders: 0, 1, 2, 3 (representing the number of people).

Dataset Creation

Curation Rationale

The dataset was created to facilitate research on WiFi-based people number estimation using low-cost ESP32-S3 devices, enabling applications in smart environments, occupancy monitoring, and crowd management.

Source Data

Initial Data Collection:

Data was collected in an indoor meeting room with a single transmitter and multiple receivers using ESP32-S3 devices. The setup included:
- Frequency Band: 2.4 GHz
- Bandwidth: 20 MHz (52 subcarriers)
- Protocol: 802.11n
- Waveform: OFDM
- Sampling Rate: ~100 Hz
- Antenna Configuration: 1 antenna per device
- Environment: Indoor with walls and a soft pad to prevent volunteer injuries.
Who are the source data producers? The data was collected by researchers, with volunteers present in a controlled meeting room environment.

Annotations

Annotation Process: Each .csv file is stored in a folder labeled with the number of people present (0–3). No additional manual annotations were provided.
Who are the annotators? The dataset creators labeled the data based on the experimental setup.

Personal and Sensitive Information

The dataset does not contain personally identifiable information, as it focuses on the number of people (0–3) without associating specific identities or biometric data beyond CSI patterns.

Citation

@ARTICLE{11278110,
  author={Zhao, Zijian and Meng, Fanyi and Lyu, Zhonghao and Li, Hang and Li, Xiaoyang and Zhu, Guangxu},
  journal={IEEE Transactions on Mobile Computing}, 
  title={CSI-BERT2: A BERT-inspired Framework for Efficient CSI Prediction and Classification in Wireless Communication and Sensing}, 
  year={2025},
  volume={},
  number={},
  pages={1-17},
  keywords={Wireless communication;Sensors;Wireless sensor networks;Predictive models;Wireless fidelity;Training;Adaptation models;Packet loss;Data models;OFDM;Channel statement information (CSI);CSI prediction;CSI classification;wireless communication;wireless sensing},
  doi={10.1109/TMC.2025.3640420}}