RadSeg is a synthetic radar dataset designed for building semantic segmentation models for radar activity recognition. Unlike existing radio classification datasets that only provide signal-wise annotations for short and isolated IQ sequences, RadSeg provides sample-wise annotations for interleaved radar pulse activities that extend across a long time horizon. This makes RadSeg the first annotated public dataset of its kind for radar activity recognition. This dataset is released to the public under the MIT License.

You can access the arXiv paper 📄 here: https://arxiv.org/abs/2312.09489

You can access the official paper 📄 here: https://ieeexplore.ieee.org/document/10445810

Z. Huang, A. Pemasiri, S. Denman, C. Fookes and T. Martin, "Multi-Stage Learning for Radar Pulse Activity Segmentation," ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Seoul, Korea, Republic of, 2024, pp. 7340-7344, doi: 10.1109/ICASSP48485.2024.10445810.

RadSeg contains pulsed radar signals at varying signal-to-noise ratios (SNRs) between -20 to 20 dB with a resolution of 0.5 dB. This repository provides the RadSeg dataset which consists of three parts:

• RadSeg-Train contains 60,000 radar signals for model training;
• RadSeg-Valadation contains 10,000 radar signals for model validation; and
• RadSeg-Test contains 10,000 radar signals held out for testing.

This dataset comprises a total of 5 radar signal types, which include:

• Barker codes, up to a code length of 13;
• Polyphase Barker codes, up to a code length of 13;
• Frank codes, up to a code length of 16;
• Linear frequency-modulated (LFM) pulses; and
• Coherent unmodulated pulse trains.

Integer mapping of class labels are given below:

# Number of signal classes
num_classes = 5

# Channel class indices, note the shape of a batch is therefore [batch_size, 6, sequence_len]
seg_classes = {
    "0": "noise", # Noise channel not used for training
    "1": "cpt",
    "2": "barker",
    "3": "polyphase",
    "4": "frank",
    "5": "lfm"
}

Additional dataset characteristics:

• The sampling rate used in RadSeg is 3.2 MHz.
• Each radar signal contains 32,768 complex, baseband IQ samples. Note, we use a subdivision of the total available samples in our ICASSP paper as part of our data augmentation workflow.
• Annotations are provided as channel-wise binary masks where each channel corresponds to a signal type.

Please refer to our conference paper for additional details on RadSeg.

To easily work with RadSeg, please feel free to use or adapt the data module in dataset.py. Below is an example workflow for creating train, validation, and test dataloaders using PyTorch:

# Root
DATA_PATH = {
    "TRAIN_DATA_PATH": "./RadSeg/train", # NOTE: Replace this as required
    "VAL_DATA_PATH": "./RadSeg/val", 
    "TEST_DATA_PATH": "./RadSeg/test"
}

# Configs
sampler=True, 
sample_window=int(4096), # Used in our ICASSP paper 
subdivision=2, # Augmentation used in our ICASSP paper
enable_mtl=False # NOTE: Not publicly available yet, but hopefully soon...

# Make train set
train_set = RadSegDataset(
    data_path=data_path["TRAIN_DATA_PATH"], 
    sampler=sampler, 
    sample_window=sample_window, 
    is_train=True,
    subdivision=subdivision,
    enable_mtl=enable_mtl
)

# Compute norm params using train set stats
train_norm_params = {
    "mu": train_set.signals_mu, # Mu is a complex number 
    "mu_real": np.real(train_set.signals_mu), 
    "mu_imag": np.imag(train_set.signals_mu),
    "var": train_set.signals_var
}

# Make validation set and test set
val_set, test_set = \
    RadSegDataset(
        data_path=data_path["VAL_DATA_PATH"], 
        sampler=sampler, 
        sample_window=sample_window, 
        is_train=False,
        normalisation_params=train_norm_params, 
        subdivision=subdivision,
        enable_mtl=enable_mtl
    ), \
    RadSegDataset(
        data_path=data_path["TEST_DATA_PATH"], 
        sampler=sampler, 
        sample_window=sample_window, 
        is_train=False,          
        normalisation_params=train_norm_params, 
        subdivision=subdivision,
        enable_mtl=enable_mtl
    )

The RadSeg datasets can be downloaded from the following links:

• RadSeg-IQ-Train - approx. file size of 29.3 GB
• RadSeg-IQ-Validation - approx. file size of 4.9 GB
• RadSeg-IQ-Test - approx. file size of 4.9 GB

• RadSeg-Masks-Train - approx. file size of 87.9 GB
• RadSeg-Masks-Validation - approx. file size of 14.6 GB
• RadSeg-Masks-Test - approx. file size of 14.6 GB

• RadSeg-SNR-Train - approx. file size of 470.8 KB
• RadSeg-SNR-Validation - approx. file size of 80.1 KB
• RadSeg-SNR-Test - approx. file size of 80.1 KB

The RadSeg dataset is published together with our paper Multi-stage Learning for Radar Pulse Activity Segmentation at the 2024 IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP 2024).

💡 Please cite both the dataset and the conference paper if you find them helpful for your research. Cheers.

@inproceedings{huang2024multi,
  title={Multi-Stage Learning for Radar Pulse Activity Segmentation},
  author={Huang, Zi and Pemasiri, Akila and Denman, Simon and Fookes, Clinton and Martin, Terrence},
  booktitle={ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages={7340--7344},
  year={2024},
  organization={IEEE}
}

💡 Our previous work may also be of interest to you:

@INPROCEEDINGS{10193318,
  author={Huang, Zi and Pemasiri, Akila and Denman, Simon and Fookes, Clinton and Martin, Terrence},
  booktitle={2023 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW)}, 
  title={Multi-Task Learning For Radar Signal Characterisation}, 
  year={2023},
  volume={},
  number={},
  pages={1-5},
  doi={10.1109/ICASSPW59220.2023.10193318}
}