In the rapidly evolving field of drone technology, ensuring secure and reliable identification of unmanned aerial vehicles (UAVs) has become a critical challenge. The proliferation of drones in both civilian and military domains has led to increased concerns over unauthorized flights, privacy violations, and potential security threats. Our research focuses on a novel detection method specifically designed for Drone-ID signals, which are integral to the communication protocols used by dominant drone manufacturers. Leveraging advanced signal processing techniques, we propose a robust framework that enhances detection probability even under extremely low signal-to-noise ratio (SNR) conditions. This article presents our work in detail, emphasizing the synergy between pulse compression and multi-scale dual correlation to achieve superior performance in drone technology applications.
The core of our approach targets the linear frequency modulation (LFM) signals that carry Zadoff-Chu (ZC) sequences within the Drone-ID packet structure of widely adopted protocols such as OcuSync 2.0. These signals are inherently energy-dispersed and vulnerable to noise, multipath fading, and frequency offsets. Traditional methods like frequency-domain cross-correlation (FDC) or fractional Fourier transform (FrFT) suffer from performance degradation at low SNR, hindering their practical deployment. Our method addresses these limitations through a three-stage architecture: signal preprocessing, multi-scale dual-correlation processing, and adaptive feature matching verification. In the following sections, we detail each stage and present comprehensive experimental results that demonstrate the efficacy of our approach.
Signal Model and Detection Principles in Drone Technology
To fully understand our detection framework, we first establish the signal model for Drone-ID transmissions. The Drone-ID data packet in OcuSync 2.0 consists of nine OFDM symbols. Among these, the fourth and sixth symbols are ZC sequences with root indices 600 and 147, respectively. These ZC sequences serve as unique identifiers for the drone. Specifically, the root index 600 ZC sequence is the primary detection target due to its fixed representation in the protocol. The ZC sequence in the frequency domain is expressed as:
$$
zc[n] = \exp\left(-j\pi u \frac{n(n+1)}{N}\right), \quad n \in [0, N-1]
$$
where \(u = 600\) is the root index, \(N = 601\) is the sequence length, and \(j\) is the imaginary unit. This sequence is modulated onto an LFM carrier, producing a time-domain signal:
$$
s(t) = A \cdot \exp\left[j(2\pi f_0 t + \pi k t^2)\right]
$$
Here, \(A\) is the amplitude, \(f_0\) the center frequency, \(k = B/T\) the chirp rate, \(T\) the pulse width, and \(B\) the bandwidth. The time-bandwidth product (TBP) \(B \cdot T\) determines the energy concentration capability. For our target LFM signal, TBP is 648, theoretically providing a processing gain of about 28.1 dB after pulse compression.
The conventional frequency-domain cross-correlation method, widely used in prior studies, computes the normalized cross-correlation coefficient between the received signal and the ideal ZC reference:
$$
\rho_{xy} = \frac{\left|\sum_{k=0}^{N-1} X_k Y_k^*\right|}{\sqrt{\sum_{k=0}^{N-1} |X_k|^2 \cdot \sum_{k=0}^{N-1} |Y_k|^2}}
$$
While this approach is straightforward, it fails to maintain high detection probability under low SNR conditions. Our proposed multi-scale dual-correlation method overcomes this by applying a two-step correlation process: first, pulse compression to focus signal energy, and second, a multi-scale delayed correlation to enhance time-frequency features. The second-order correlation output is given by:
$$
C_2(\tau, t) = \sum_{m=1}^{M} \alpha_m \cdot \left[ y(t) \cdot y^*(t + \tau_m) \right]
$$
where \(\tau_m\) represents the delay parameter for scale \(m\), \(\alpha_m\) is the weighting coefficient, and \(M\) is the total number of scales. This multi-scale fusion captures both coarse and fine temporal structures, effectively suppressing random noise while reinforcing the target signal’s correlation peaks.
Multi-Scale Dual-Correlation Detection Algorithm
Our algorithm is structured into three cascaded stages:
Stage 1: Signal Preprocessing for Drone Technology Robustness
In real-world scenarios, carrier frequency offset (CFO) caused by relative motion and oscillator drifts can severely degrade correlation performance. We implement an adaptive CFO compensation method based on correlation peak maximization. The received signal model with CFO \(\Delta f\) is:
$$
r(t) = s(t) \cdot \exp(j2\pi \Delta f t) + n(t)
$$
We search over a candidate range \(\Delta f \in [-0.1B, 0.1B]\) (where \(B\) is signal bandwidth) and select the value that maximizes the correlation between the compensated signal and the reference LFM template. This greedy search ensures minimal residual CFO. Additionally, we apply an adaptive Butterworth bandpass filter to suppress out-of-band noise. The filter order \(n\) is dynamically adjusted based on estimated SNR: \(n=8\) for SNR < -3 dB, and \(n=4\) for SNR > 0 dB. This preprocessing step yields a typical SNR improvement of 3–5 dB, providing a cleaner input for subsequent stages.
Stage 2: First Correlation – Pulse Compression
Pulse compression is realized via matched filtering. The matched filter kernel \(h(t)\) is the time-reversed conjugate of the transmitted LFM signal:
$$
h(t) = s^*(T_0 – t)
$$
For computational efficiency, we implement the convolution in the frequency domain using FFT:
$$
Y[k] = X[k] \cdot S_{\text{ref}}^*[k]
$$
followed by inverse FFT to obtain the time-domain compressed output \(C_1[n]\). The peak amplitude after compression scales with the TBP, achieving the theoretical processing gain. However, under low SNR, noise-induced spurious peaks can obscure the genuine target peak. This motivates the need for a second correlation stage.
Stage 3: Second Correlation – Multi-Scale Enhancement
We design \(M = 15\) scales with delay parameters \(\tau_m = \beta_m \cdot (T \cdot F_s / 2)\), where \(\beta_m\) ranges from 0.1 to 1.5 in steps of 0.1. For each scale, we compute:
$$
C_{2,m}[n] = C_1[n] \cdot \text{conj}(C_1[n + \tau_m])
$$
These outputs are then combined using adaptive weights based on the SNR of each scale:
$$
C_2[n] = \sum_{m=1}^{M} w_m \cdot |C_{2,m}[n]|, \quad w_m = \frac{\text{SNR}_m}{\sum_{i=1}^{M} \text{SNR}_i}
$$
where \(\text{SNR}_m = (\mu_{C_{2,m}} – \mu_{\text{bg}})/\sigma_{\text{bg}}\) is computed from the mean and standard deviation of background noise. This fusion maximizes the contribution of scales with higher signal-to-noise ratios, effectively eliminating random noise peaks while preserving structured target features.
Adaptive Feature Matching Verification
After obtaining the enhanced correlation output \(C_2[n]\), we extract candidate peaks exceeding a threshold \(\gamma = \mu_{\text{bg}} + 3\sigma_{\text{bg}}\). For each candidate, we compute three features:
- Sharpness: \( \text{Sharpness} = (P_{\text{peak}} – \mu_{\text{bg}})/\sigma_{\text{bg}} \)
- Symmetry: \( \text{Symmetry} = 1 – ||L – \text{flip}(R)|| / (||L|| + ||R||) \)
- Width Score: \( \text{WidthScore} = \exp(-|FWHM – W_0|/W_0) \)
The overall score is computed as:
$$
\text{Score} = 0.4 \cdot \text{Sharpness} + 0.3 \cdot \text{Symmetry} + 0.3 \cdot \text{WidthScore}
$$
The optimal weights (0.4, 0.3, 0.3) are obtained via a 5-fold cross-validation search over a grid of weight combinations. The adaptive threshold is set as:
$$
\theta_{\text{score}} = \theta_0 – 0.05 \cdot \text{SNR}_{\text{in}}
$$
with \(\theta_0 = 0.6 – 0.02 \cdot (\mu_{\text{bg}}/\sigma_{\text{bg}})\). If the candidate’s score exceeds the threshold, it is classified as a genuine Drone-ID signal; otherwise, it is rejected as a false peak.
Experimental Results and Performance Evaluation
Experimental Setup
We evaluate our method using both real-world captured signals from DJI mini2, mini3, and mavic air drones, and simulated signals generated according to OcuSync 2.0 parameters (\(T=72~\mu\text{s}\), \(B=9~\text{MHz}\), \(N=601\), \(u=600\)). The SNR range from -10 dB to +10 dB is considered with additive white Gaussian noise and random CFO up to \(\pm 10\%\) of bandwidth. Monte Carlo simulations with over 500 iterations are performed to ensure statistical stability. The following table summarizes the key parameters.
| Parameter | Value |
|---|---|
| Pulse Width \(T\) | 72 \(\mu\)s |
| Bandwidth \(B\) | 9 MHz |
| ZC Sequence Length \(N\) | 601 |
| Root Index \(u\) | 600 |
| Number of Scales \(M\) | 15 |
| Monte Carlo Iterations | 500 |
| SNR Range | -10 dB to +10 dB |
| CFO Range | \(\pm 0.9\) MHz |
Detection Probability vs. False Alarm Rate
Figure 1 (embedded above) illustrates the ROC curves comparing our multi-scale dual-correlation (MSDC) method against FDC, FrFT, and WHT. At SNR = -5 dB, MSDC achieves a detection probability of 91.2%, surpassing FDC (60.1%), FrFT (36.1%), and WHT (25.5%). The false alarm rate for MSDC remains below 1% across all tested SNR levels, while FrFT shows uncontrollable false alarms even at high SNR. This demonstrates the superior noise suppression capability of our dual-correlation architecture.
Precision-Recall Analysis
The precision-recall curves reveal that MSDC maintains near-ideal balance (precision 0.947, recall 1.0) across a wide SNR range. In contrast, FDC degrades rapidly below 0 dB SNR, and FrFT/WHT consistently underperform even at high SNR. The adaptive feature verification mechanism ensures that false positives are minimized without sacrificing true positive detections.
Computational Efficiency
We measure per-sample processing time on an Intel i5-10500K platform. The following table compares the computational load of each method.
| Method | Average Time (ms) | Relative Complexity |
|---|---|---|
| FDC | 12 | \(O(N \log N)\) |
| FrFT | 68 | \(O(N^2)\) |
| WHT | 8 | \(O(N \log N)\) |
| MSDC (Proposed) | 35 | \(O(N \log N)\) |
Although MSDC is slower than FDC and WHT due to multi-scale fusion, its processing time (35 ms) is well within the real-time requirement (500 ms per sample). For hardware deployment on SDR platforms (e.g., AD9361 + Zynq-7000 FPGA), the algorithm can be further optimized using parallel IP cores, reducing latency to under 500 ms.
Radar Chart of Comprehensive Performance
| Metric | FDC | FrFT | WHT | MSDC |
|---|---|---|---|---|
| Detection Probability | 0.60 | 0.36 | 0.26 | 0.91 |
| Precision | 0.55 | 0.30 | 0.22 | 0.95 |
| Recall | 0.62 | 0.38 | 0.28 | 0.94 |
| False Alarm Suppression | 0.88 | 0.70 | 0.80 | 0.99 |
| Computational Efficiency | 0.92 | 0.15 | 0.95 | 0.50 |
From the radar chart (conceptually combining these metrics), MSDC dominates in detection-oriented dimensions while maintaining acceptable computational efficiency.
Discussion and Conclusions
The proposed multi-scale dual-correlation method represents a significant advancement in drone technology for reliable Drone-ID signal detection. By cascading pulse compression with an adaptive multi-scale correlation stage, we achieve high detection probability and low false alarm rate even under challenging low-SNR conditions. The adaptive feature matching further ensures robustness against spurious peaks. Our experiments confirm that at SNR = -5 dB, the detection probability reaches 91.2%, outperforming FDC by 31.1 percentage points. The false alarm rate stays below 1%, which is critical for practical deployment in drone technology security systems.
Future work will focus on extending the method to handle high-speed drone maneuvers (velocity > 50 m/s) and further reducing hardware resource consumption for real-time embedded implementations. The integration of this algorithm into software-defined radio platforms for field trials will validate its applicability in real-world drone technology monitoring scenarios.