A manifold learning approach for gesture recognition from micro-Doppler radar measurements

E. S. Mason, H. N. Mhaskar, A. Guo (2022)

doi.org/10.1016/j.neunet.2022.04.024

Data and setup

We use the DopNet radar dataset, which can be found on their GitHub repository.

1. Download data from persons A to F (Data_Per_PersonData_Training_Person_{A-F}.mat) and put those six files into the data/dopnet/ directory
2. Install required Python libraries (e.g. in a Conda environment): pip install -r requirements.txt
3. Install PyTorch
4. Generate .npy files: python export_data.py
5. The example_usage.py script demonstrates how to import and preprocess the dataset as well as run the models. The script itself can be run: python example_usage.py

Experiments

To run the experiments presented in the paper: cd tests and python experiments.py. The following five experiments are available:

1. Unused (replaced by experiment 3)
2. Evaluate each model by training on 5 subjects' data and testing on the last subject's data
3. Evaluate each model on different train/test split ratios
4. Evaluate PCA-based models across PCA dimensions
5. Plot singular values of spectrograms

The models used are referred to by the following names, corresponding to Tables 3 and 4 of the paper:

• SVD-based models: Gaussian SVD SVM, Grassmann SVD SVM, Laplace SVD SVM
• PCA-based models: PCA KNN, PCA LocSVM16, PCA LocSVM64
• CNN models: CNN1, CNN2

For reference, the experiments were run on Arch Linux (kernel 5.x) using an AMD Ryzen 5 5600X with 32GB of memory. An NVIDIA GTX 1060 6GB was used to accelerate PyTorch for CNN training and inference. The experiments may need modifications to run on less than 32GB of memory or without an NVIDIA GPU.

File rundown

• src/manifold_svm.py
  • Contains all the classification code. Classifiers inherit from sklearn.SVM.SVC, and only change the kernel function
• src/preprocessing.py
  • Contains a number of preprocessing classes designed to be used in sklearn.Pipeline
  • Only Threshold is used, which implements scaling to binary/[-1, 0] interval and Otsu/Yen thresholding
• export_data.py
  • Run this to read and export DopNet data
  • Put all the Data_Per_PersonData....mat files into data/dopnet/
• tests/dopnet.py
  • Contains wrapper functions for import/exporting data and running trials
  • Trials are outdated, mostly replaced by tests/experiments.py
• tests/experiments.py
  • Contains all experiments run for this paper with helper functions for exporting test results
• tests/models.py
  • Contains modeling code, repackaged from src/manifold_svm.py for ease of use in experiments
• tests/plotting.py
  • Contains helper code for plotting experimental results