RadDet: A Wideband Dataset for Real-Time Radar Spectrum Detection

Zi Huang, Simon Denman, Akila Pemasiri, Terrence Martin, Clinton FookesICASSP 2025 · OpenReview

TL;DR

This paper introduces RadDet, a novel and large-scale dataset for real-time wideband radar spectrum detection, addressing the critical scarcity of such data. They also establish a benchmark by evaluating state-of-the-art object detection models on this dataset, revealing trade-offs between detection accuracy and inference speed.

What Problem Does This Paper Address?

In electronic warfare (EW), accurately detecting and identifying radar signals in a complex, wideband radio frequency spectrum is crucial for situational awareness. This allows for the timely deployment of countermeasures. However, developing and evaluating effective detection models is severely hampered by the lack of publicly available, diverse, and well-annotated datasets of wideband radar signals. Existing datasets are often limited in bandwidth, signal variety, or lack the necessary time-frequency annotations for detection tasks. This paper aims to bridge this gap by creating a comprehensive dataset and establishing a benchmark for this challenging problem.

Methods

The core contribution is the creation of the RadDet dataset, a synthetic collection of wideband radar signals. The dataset is generated with a focus on realistic EW scenarios, including:

  • Wideband Spectrum: Signals span a 500 MHz frequency band.
  • Diverse Radar Classes: 11 distinct radar signal types are included, encompassing pulsed, linear frequency-modulated (LFM), and advanced LPI (Low Probability of Intercept) polyphase codes.
  • Varied SNR: Signals are generated across 6 different Signal-to-Noise Ratio (SNR) settings, simulating degraded environments.
  • Radar Density: Two environments are simulated: sparse (at most one radar per frame) and dense (up to 9 radars per frame), reflecting different operational scenarios.
  • Time-Frequency Resolution: Data is provided at 3 different time-frequency resolutions.
  • Annotations: Each frame (1 million I/Q samples) is accompanied by time-frequency bounding boxes and class labels, formatted for object detection.

To establish a benchmark, the authors adapt standard computer vision object detection models (YOLOv3, YOLOv6, YOLOv9, and RT-DETR) for radar spectrum detection. They convert the raw I/Q data into spectrograms using the Short-Time Fourier Transform (STFT) and employ a max-hold spectrogram approach for efficient representation of wideband signals.

Datasets

  • NIST-CBRS 3.5 GHz Dataset: A synthetic dataset from NIST, used as a baseline. It contains 5 radar classes across a 10 MHz bandwidth and limited SNR range. The authors modified it to provide time-frequency annotations for detection.
  • RadDet Dataset: The primary contribution of this paper. It's a new, large-scale, open-source dataset comprising 40,000 radar frames. It features 11 radar classes, 6 SNR settings, 2 radar density environments, and 3 time-frequency resolutions, with comprehensive time-frequency and class annotations. RadDet is split into training (20,000 frames), validation (14,000 frames), and test (6,000 frames) sets.

Key Insights and Ideas

  • Data is King (Especially for Radar): The paper strongly emphasizes that the scarcity of wideband radar datasets is a major bottleneck for progress in EW. RadDet aims to alleviate this.
  • Object Detection for Spectrum: Applying established object detection frameworks from computer vision to the radar spectrum detection problem is a viable and effective approach.
  • Trade-offs in Time-Frequency Resolution: Higher time-frequency resolution leads to better detection accuracy (mAP) but significantly reduces inference speed (FPSavg). This highlights a critical design choice for real-time systems.
  • Data Augmentation is Crucial: Augmenting the training data (translations, scaling, etc.) significantly boosts detection performance across all tested models.
  • Sparse Environments are Tough: Detecting signals in sparse environments (RadDet-1T) is more challenging due to infrequent pulses and smaller object footprints compared to denser scenarios or the NIST-CBRS dataset.
  • Model Suitability Varies: While YOLO family models generally perform well, RT-DETR, despite its strengths in natural image tasks, is less suited for radar spectrum detection, potentially due to its reliance on pre-trained natural image embeddings.

Key Results

  • Benchmark Performance: The paper provides a comprehensive performance comparison of YOLOv3, YOLOv6, YOLOv9, and RT-DETR across different sizes and datasets.
  • Accuracy vs. Speed: For example, reducing the input size of YOLOv3 from M (256x256) to S (128x128) resulted in a 49% speedup but a 7% and 39% drop in mAP50 on NIST-CBRS and RadDet-9T, respectively.
  • Data Augmentation Impact: Models trained with data augmentation consistently outperformed those without, showing significant improvements in mAP.
  • Dense vs. Sparse: Performance metrics (mAP) were generally lower on the sparse RadDet-1T dataset compared to the dense RadDet-9T and NIST-CBRS datasets, especially for certain models.

Limitations and Future Work

  • Synthetic Data: While comprehensive, RadDet is a synthetic dataset. Future work could involve incorporating real-world radar data to further validate model performance.
  • Model Architectures: The paper focuses on established object detection models. Exploring novel architectures specifically designed for radar signal characteristics could yield further improvements.
  • Real-time Constraints: While evaluating inference speed, the paper acknowledges that further optimization might be needed for deployment on resource-constrained EW platforms.
  • Beyond Detection: Future research could extend to more complex tasks like radar signal classification, parameter estimation, and tracking within the RadDet framework.

Summary updated

Abstract

Real-time detection of radar signals in a wideband radio frequency spectrum is a critical situational assessment function in electronic warfare. Compute-efficient detection models have shown great promise in recent years, providing an opportunity to tackle the spectrum detection problem. However, progress in radar spectrum detection is limited by the scarcity of publicly available wideband radar signal datasets accompanied by corresponding annotations. To address this challenge, we introduce a novel and challenging dataset for radar detection (RadDet), comprising a large corpus of radar signals occupying a wideband spectrum across diverse radar density environments and signal-to-noise ratio (SNR) settings. RadDet contains 40,000 frames, each generated from 1 million in-phase and quadrature (I/Q) samples across a 500 MHz frequency band. RadDet includes 11 classes of radar signals across 6 different SNR settings, 2 radar density environments, and 3 different time-frequency resolutions, with corresponding time-frequency and class annotations. We evaluate the performance of state-of-the-art real-time detection models on RadDet and a modified radar classification dataset from NIST (NIST-CBRS) to establish a novel benchmark for wideband radar spectrum detection.
BibTeX
@inproceedings{huang2025raddet,
  author = {Zi Huang and Simon Denman and Akila Pemasiri and Terrence Martin and Clinton Fookes},
  title = {{RadDet: A Wideband Dataset for Real-Time Radar Spectrum Detection}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2025},
  url = {https://openreview.net/forum?id=t3NkHrAB9y},
}
Figure