WiHAR: Privacy-Preserving Human Sensing using WiFi CSI

---

This project implements human activity recognition using WiFi signals instead of cameras. It uses Channel State Information (CSI), turns it into images, and classifies activities with a Vision Transformer (ViT) model. The project report is available here.

Setup

1. Install Python and dependencies:

conda create -n wihar python=3.9
conda activate wihar
pip install -r requirements.txt

2. Download NTU-Fi HAR CSI files and place them in:

data/NTU-Fi_HAR/train_amp/<activity_name>/*.mat
data/NTU-Fi_HAR/test_amp/<activity_name>/*.mat

Usage

1. Generate images from CSI:

python cfr_out.py

• This creates images in data/cfr_dataset/train/ and data/cfr_dataset/test/.

2. Create a validation split by copying a portion (10–20%) of images from each class in train/ to a new val/ folder:

data/cfr_dataset/
  ├── train/
  ├── val/
  └── test/

3. Train the model:

python vit_train.py fit \
  --data.dataset custom \
  --data.root data/cfr_dataset/ \
  --data.num_classes 6 \
  --data.batch_size 32 \
  --trainer.max_steps 1800 \
  --trainer.check_val_every_n_epoch 2 \
  --model.warmup_steps 180 \
  --model.lr 0.01

4. Test the model:

python vit_train.py test \
  --ckpt_path <path_to_best_checkpoint> \
  --data.dataset custom \
  --data.root data/cfr_dataset/ \
  --data.num_classes 6

• Replace <path_to_best_checkpoint> with the actual checkpoint file.

Acknowledgements

Thanks to SenseFi and vit-finetune.