RFUAV
This is official repository of our paper "RFUAV: A Benchmark Dataset for Unmanned Aerial Vehicle Detection and Identification". Codes include a two stage model to achieve drone detection and classification using some FFT/STFT analytical method. The Raw data will be free to use after our paper is accept. Star us!!!!, if you think this is useful♥
Install / Use
/learn @kitoweeknd/RFUAVREADME
Abstract
The official repository for our paper, "RFUAV: A Benchmark Dataset for Unmanned Aerial Vehicle Detection and Identification", can be accessed here. RFUAV offers a comprehensive benchmark dataset for Radio-Frequency (RF)-based drone detection and identification.
In addition to the dataset, we provide the raw data used to generate the spectral information, which includes recordings from 35 different types of drones under high signal-to-noise ratio (SNR) conditions. This dataset is available to all researchers working with RF data for drone analysis. Researchers can apply the deep learning methods we have provided, or use traditional signal processing techniques such as decoding, demodulation, and FFT.
Detailed information about the dataset, including file sizes (total data volume for each drone), SNR (the highest SNR for each dataset), and the middle frequency (the central frequency used during data collection for each drone), is provided in the figure below.
We analyzed the properties of each drone in the dataset, including Frequency Hopping Signal Bandwidth (FHSBW), Frequency Hopping Signal Duration Time (FHSDT), Video Transmitted Signal Bandwidth (VSBW), Frequency Hopping Signal Duty Cycle (FHSDC), and Frequency Hopping Signal Pattern Period (FHSPP). The distributions of these properties are plotted below. More detailed information can be found in our paper.
With RFUAV, you can achieve drone signal detection and identification directly on raw IQ data, as demonstrated below:
<div style="text-align:center;">
</div>
1. Quick Start
<details> <summary>Installation</summary>pip install -r requirements.txt
python inference.py
python train.py
2. Usage
SDR Playback
Since our data was directly collected using USRP devices, it is fully compatible with USRP and GNU Radio for signal replay. You can use our raw data to broadcast signals through radio equipment to achieve your desired outcomes. Additionally, we provide the replay results observed during our experiments using an oscilloscope for reference.
<div style="text-align:center;">
</div>
2.1 Converting Raw Frequency Signal Data to Spectrograms
Python Pipeline
We provide a signal processing pipeline to convert binary raw frequency signal data into spectrogram format using both MATLAB and Python toolboxes.
Visualize Spectrograms
You can easily visualize the spectrogram of a specific data pack using the following code. The oneside parameter controls whether to display the half-plane or full-plane spectrogram.
from graphic.RawDataProcessor import RawDataProcessor
datapack = 'Your datapack path'
test = RawDataProcessor()
test.ShowSpectrogram(data_path=datapack,
drone_name='DJ FPV COMBO',
sample_rate=100e6,
stft_point=2048,
duration_time=0.1,
oneside=False,
Middle_Frequency=2400e6)
Batch Conversion to Images
To automatically convert raw frequency signal data into spectrograms and save them as PNG images:
from graphic.RawDataProcessor import RawDataProcessor
data_path = 'Your datapack path'
save_path = 'Your save path'
test = RawDataProcessor()
test.TransRawDataintoSpectrogram(fig_save_path=save_path,
data_path=data_path,
sample_rate=100e6,
stft_point=1024,
duration_time=0.1)
Save as Video
You can use the TransRawDataintoVideo() method to save the spectrogram as a video, which provides better visualization of temporal signal evolution:
from graphic.RawDataProcessor import RawDataProcessor
data_path = 'Your datapack path'
save_path = 'Your save path'
test = RawDataProcessor()
test.TransRawDataintoVideo(save_path=save_path,
data_path=data_path,
sample_rate=100e6,
stft_point=1024,
duration_time=0.1,
fps=5)
Waterfall Spectrogram
The waterfall_spectrogram() function converts raw data into a waterfall spectrogram video, visually displaying how the signal evolves over time:
from graphic.RawDataProcessor import waterfall_spectrogram
datapack = 'Your datapack path'
save_path = 'Your save path'
images = waterfall_spectrogram(datapack=datapack,
fft_size=256,
fs=100e6,
location='buffer',
time_scale=39062)
MATLAB Pipeline
You can use the check.m program to visualize the spectrogram of a specific data pack:
data_path = 'Your datapack path';
nfft = 512;
fs = 100e6;
duration_time = 0.1;
datatype = 'float32';
check(data_path, nfft, fs, duration_time, datatype);
2.2 SNR Estimation and Adjustment
We provide SNR estimation and adjustment tools using the MATLAB toolbox to help you analyze and process binary raw frequency signal data.
SNR Estimation
First, locate the signal position and estimate the SNR:
[idx1, idx2, idx3, idx4, f1, f2] = positionFind(dataIQ, fs, bw, NFFT);
snr_esti = snrEsti(dataIQ, fs, NFFT, f1, f2, idx1, idx2, idx3, idx4);
SNR Adjustment
The awgn1 function adjusts the noise level of raw signal data based on the SNR estimation results. The signal-to-noise ratio can be adjusted between -20 dB and 20 dB, with a default step size of 2 dB. You can also define a custom scale if needed.
2.3 Training Custom Drone Classification Models
We provide custom training code for drone identification tasks based on the PyTorch framework. Currently supported models include ViT, ResNet, MobileNet, Swin Transformer, EfficientNet, DenseNet, VGG, and many others. You can also customize your own model using the code in utils.trainer.model_init_().
Training
To customize the training, create or modify a configuration file with the .yaml extension and specify its path in the training code. You can adjust the arguments in utils.trainer.CustomTrainer() to achieve the desired training setup:
from utils.trainer import CustomTrainer
trainer = CustomTrainer(cfg='Your configuration file path')
trainer.train()
Alternatively, you can use the base trainer directly:
from utils.trainer import Basetrainer
trainer = Basetrainer(
model='resnet50',
train_path='Your train data path',
val_path='Your val data path',
num_class=23,
save_path='Your save path',
weight_path='Your weights path',
device='cuda:0',
batch_size=32,
shuffle=True,
image_size=224,
lr=0.0001
)
trainer.train(num_epochs=100)
Inference
We provide an inference pipeline that allows you to run inference on either spectrogram images or binary raw frequency data. When processing binary raw frequency data, the results are automatically packaged into a video with identification results displayed on the spectrogram. Note: When inferring on binary raw frequency data, you must use a model weight trained on the spectrogram dataset.
from utils.benchmark import Classify_Model
test = Classify_Model(cfg='Your configuration file path',
weight_path='Your weights path')
test.inference(source='Your target data path',
save_path='Your target save path')
2.4 Training Custom Drone Detection Models
We provide custom training methods for drone detection tasks. Currently supported models include YOLOv5.
Training
You can train the YOLOv5 model for drone detection using the following code:
from utils.trainer import DetTrainer
model = DetTrainer(cfg='Your configuration file path', dataset_dir = "Your dataset file path")
model.train()
Inference
The inference pipeline allows you to run your model on either spectrogram images or binary raw frequency data. When processing binary raw frequency data, the results are automatically packaged into a video with detection results displayed on the spectrogram. Note: When inferring on binary raw frequency data, you must use a model weight trained on the spectrogram dataset.
from utils.benchmark import Detection_Model
test = Detection_Model(cfg='Your configuration file path',
weight_path='Your weights path')
test.inference(source='Your target data path',
save_dir='Your target save path')
2.5 Two-Stage Detection and Classification
We provide a two-stage pipeline that combines detection and classification: the first stage detects drone signals, and the second stage classifies the detected signals. You can process raw data packs directly, and the results w
