In recent years, the proliferation of China UAV (Unmanned Aerial Vehicle) technology has brought tremendous benefits to sectors such as aerial surveying, logistics, and agriculture. However, the misuse of China UAVs, including unauthorized flights in restricted airspace and potential security threats, has raised serious concerns. Among the global UAV market, DJI, a leading China UAV manufacturer, dominates with over 85% market share. Its proprietary OcuSync protocol, particularly the Drone-ID signal, embeds Zadoff-Chu (ZC) sequences that serve as synchronization pilots. Detecting these Drone-ID signals reliably under low signal-to-noise ratio (SNR) conditions is critical for China UAV regulation and safety. Traditional detection methods, such as frequency domain cross-correlation (FDC), fractional Fourier transform (FrFT), and Walsh-Hadamard transform (WHT), exhibit significant performance degradation in low SNR and frequency offset scenarios. To address these challenges, we propose a novel detection method based on pulse compression and multi-scale dual correlation. This approach enhances signal energy focusing through matched filtering and further strengthens time-frequency features via dual correlation, followed by an adaptive feature matching verification mechanism. Extensive experiments demonstrate that our method achieves a detection probability of 91.2% at an SNR of -5 dB, outperforming FDC by 31.1 percentage points, while maintaining a false alarm rate below 1%. Our work provides a robust and practical solution for China UAV Drone-ID signal detection in complex electromagnetic environments.
Introduction
The rapid adoption of China UAV has revolutionized numerous industries, but it has also introduced new challenges in airspace management and public safety. The ability to detect and identify China UAVs in real time is essential for mitigating risks such as illegal surveillance, smuggling, and potential collisions. DJI’s OcuSync protocol, widely used in China UAV communication, incorporates Drone-ID packets that contain predefined ZC sequences—specifically with root indices 600 and 147. These sequences are modulated as linear frequency modulation (LFM) signals, which are inherently wideband and energy-dispersive, making them suitable for pulse compression techniques. However, practical scenarios often involve severe noise, multipath fading, and Doppler shifts, leading to SNRs below 0 dB. Conventional detection algorithms, while effective in high SNR, fail under these conditions. In this paper, we present a multi-scale dual correlation detection framework that integrates pulse compression for initial energy focusing, followed by a secondary correlation across multiple time delays to enhance signal features, and finally an adaptive verification stage to discriminate true signals from noise peaks. Our contributions include: (1) a novel dual correlation architecture that improves detection robustness against frequency offsets and interference; (2) a multi-dimensional feature extraction and adaptive thresholding scheme; (3) comprehensive performance evaluation using both real-world collected China UAV signals and synthetic data.
Signal Model and Detection Principles
OcuSync Protocol Signal Model
The Drone-ID signal in DJI’s OcuSync 2.0 protocol consists of 9 OFDM symbols, with the 4th and 6th symbols being ZC sequences of length 601 and root indices 600 and 147, respectively. The ZC sequence is expressed in the frequency domain 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, and \( N = 601 \) corresponds to the number of active subcarriers plus guard band. This sequence is transmitted via an LFM signal, which in continuous time is given by:
$$
s(t) = A \cdot \exp\left[j\left(2\pi f_0 t + \pi k t^2\right)\right]
$$
where \( A \) is the amplitude, \( f_0 \) is the center frequency, \( k = B/T \) is the chirp rate, \( T \) is the pulse width, and \( B \) is the bandwidth. The transmitted LFM signal possesses a large time-bandwidth product (TBP = B·T), which enables significant pulse compression gain and inherent interference resilience.
Traditional Frequency Domain Cross-Correlation
The FDC method computes the normalized cross-correlation coefficient between the received signal and a reference ZC sequence in the frequency domain:
$$
\rho_{xy} = \frac{\left| \sum_{k=0}^{N-1} X_k Y_k^* \right|}{\sqrt{\sum_{k=0}^{N-1} |X_k|^2 \cdot \sum_{k=0}^{N-1} |Y_k|^2}}
$$
Although straightforward, FDC suffers from severe performance degradation under low SNR and frequency offset conditions due to its inability to combat noise and interference.
Multi-Scale Dual Correlation Detection Principle
Our proposed multi-scale dual correlation method first applies pulse compression to focus the signal energy:
$$
y(t) = s(t) * h(t)
$$
where \( h(t) \) is the matched filter kernel. Then, a second correlation is performed with multiple delay scales to exploit temporal structure:
$$
C_2(\tau, t) = \sum_{m=1}^{M} \alpha_m \cdot \left[ y(t) \cdot y^*(t + \tau_m) \right]
$$
Here, \( \tau_m \) are delay parameters and \( \alpha_m \) are weighting coefficients. This dual processing enhances the target peak while suppressing random noise peaks, leading to robust detection even at very low SNR.
Algorithm Framework
Signal Preprocessing
Before detection, we perform two critical preprocessing steps. First, adaptive frequency offset compensation is applied to correct carrier frequency offsets (CFO) caused by relative motion between the China UAV and the receiver. The compensation is performed by searching over a range \( [\Delta f_{\min}, \Delta f_{\max}] = [\pm 10\% B] \) and selecting the offset that maximizes the correlation peak:
$$
r'(t) = r(t) \exp(-j 2\pi \Delta \hat{f} t)
$$
Second, an adaptive Butterworth bandpass filter is employed to suppress out-of-band noise. The filter order \( n \) is adjusted based on the estimated SNR: for SNR below -3 dB, \( n=8 \); for SNR above 0 dB, \( n=4 \). This preprocessing improves the SNR by 3–5 dB on average, providing a cleaner input for subsequent correlation stages.
First Correlation: Pulse Compression
We implement pulse compression using fast convolution in the frequency domain to reduce computational complexity to \( O(N \log N) \). The matched filter kernel is derived from the reference LFM signal, and the output is:
$$
C_1[n] = \text{IFFT}\left\{ \text{FFT}\{x[n]\} \cdot \text{conj}(\text{FFT}\{s_{\text{ref}}[n]\}) \right\}
$$
The pulse compression gain theoretically reaches \( 10 \log_{10}(TB) \). For our parameters ( \( T=72\,\mu s, B=9\,\text{MHz} \) ), the gain is about 28.1 dB, which is crucial for detecting weak signals from China UAV at long ranges.
Second Correlation: Multi-Scale Enhancement
The output of pulse compression \( C_1[n] \) may still contain false peaks from noise. To further enhance genuine target features, we perform a second correlation with 15 delay scales (\( M=15 \)) ranging from 0.1 to 1.5 in steps of 0.1, normalized by the peak half-width. For each scale \( m \):
$$
C_{2,m}[n] = C_1[n] \cdot \text{conj}(C_1[n + \tau_m])
$$
These scale-specific outputs are then fused using SNR-adaptive weights:
$$
C_2[n] = \sum_{m=1}^{M} w_m \cdot |C_{2,m}[n]|, \quad w_m = \frac{SNR_m}{\sum_{i=1}^{M} SNR_i}
$$
The SNR for each scale is computed as \( SNR_m = (\mu_{C_{2,m}} – \mu_{bg}) / \sigma_{bg} \). This fusion amplifies consistent structural patterns across scales while suppressing random fluctuations.
Adaptive Feature Matching Verification
After obtaining \( C_2[n] \), candidate peaks above a threshold \( \gamma = \mu_{bg} + 3\sigma_{bg} \) are evaluated using three features:
- Sharpness: \( \text{Sharpness} = (P_{\text{peak}} – \mu_{bg}) / \sigma_{bg} \)
- Symmetry: \( \text{Symmetry} = 1 – \|L – \text{flip}(R)\| / (\|L\| + \|R\|) \)
- Width Score: \( \text{WidthScore} = \exp(-|\text{FWHM} – W_0| / W_0) \)
A composite score is computed as:
$$
\text{Score} = \alpha_1 \cdot \text{Sharpness} + \alpha_2 \cdot \text{Symmetry} + \alpha_3 \cdot \text{WidthScore}
$$
with optimal weights \( (\alpha_1, \alpha_2, \alpha_3) = (0.4, 0.3, 0.3) \) obtained via 5-fold cross-validation. The decision threshold adapts to the input SNR:
$$
\theta_{\text{score}} = \theta_0 – \kappa \cdot SNR_{\text{in}}, \quad \theta_0 = 0.6 – 0.02 \cdot \frac{\mu_{bg}}{\sigma_{bg}}, \quad \kappa = 0.05
$$
Candidates with Score ≥ \(\theta_{\text{score}}\) are declared as genuine China UAV signals; otherwise, they are rejected. An iterative adjustment mechanism further refines the threshold when no candidate is found, ensuring robust operation across varying conditions.
Experimental Results and Analysis
Real-World Signal Experiment
We collected Drone-ID signals from multiple DJI China UAV models (Mini 2, Mini 3, Mavic Air) in a laboratory environment with non-ideal conditions (SNR ≈ -3 dB). After pulse compression, a noticeable peak emerged at the symbol position corresponding to root index 600. After the second dual correlation, the normalized peak amplitude reached above 0.9, with significantly reduced false peaks. This validates the effectiveness of our two-stage enhancement in practical scenarios involving China UAV signals.
Synthetic Data Performance Comparison
We generated synthetic test sets with SNRs ranging from -10 dB to 10 dB, adding Gaussian noise and random frequency offsets up to ±10% of bandwidth. Monte Carlo simulations were performed with 500+ iterations. The comparison methods include FDC, FrFT, WHT, and our proposed multi-scale dual correlation (MSDC).
Detection Probability and False Alarm Rate
Table I summarizes the detection probability (\( P_d \)) and false alarm rate (\( P_{fa} \)) for different methods at various SNRs.
| SNR (dB) | Method | \( P_d \) | \( P_{fa} \) |
|---|---|---|---|
| -5 | FDC | 0.601 | 0.048 |
| -5 | FrFT | 0.361 | 0.172 |
| -5 | WHT | 0.255 | 0.213 |
| -5 | MSDC | 0.912 | 0.009 |
| 0 | FDC | 0.823 | 0.021 |
| 0 | FrFT | 0.501 | 0.096 |
| 0 | WHT | 0.402 | 0.144 |
| 0 | MSDC | 0.974 | 0.003 |
| 10 | FDC | 0.992 | 0.005 |
| 10 | FrFT | 0.778 | 0.038 |
| 10 | WHT | 0.611 | 0.087 |
| 10 | MSDC | 0.998 | 0.001 |
At SNR = -5 dB, MSDC achieves a detection probability of 91.2%, which is 31.1 percentage points higher than FDC, 55.1% higher than FrFT, and 65.7% higher than WHT. The false alarm rate of MSDC remains below 1% even at this challenging SNR, while the other methods show significantly higher false alarms, especially FrFT and WHT. These results confirm the superior noise suppression and peak enhancement capabilities of our dual correlation framework for China UAV signal detection.
Precision-Recall Analysis
Table II presents the precision and recall values at the optimal operating points for each method at SNR = -5 dB and 0 dB.
| SNR (dB) | Method | Precision | Recall |
|---|---|---|---|
| -5 | FDC | 0.785 | 0.601 |
| -5 | FrFT | 0.512 | 0.355 |
| -5 | WHT | 0.389 | 0.231 |
| -5 | MSDC | 0.947 | 1.000 |
| 0 | FDC | 0.916 | 0.823 |
| 0 | FrFT | 0.675 | 0.501 |
| 0 | WHT | 0.523 | 0.402 |
| 0 | MSDC | 0.988 | 0.974 |
MSDC achieves the best balance, with a precision of 0.947 and recall of 1.0 at -5 dB, meaning it can detect all genuine China UAV signals without missing any, while maintaining very few false positives. In contrast, FDC sacrifices recall for precision, and FrFT and WHT suffer from both low precision and low recall, making them unreliable for practical China UAV monitoring.
Computational Efficiency
Table III compares the average single-sample computation time (in milliseconds) for each method on an Intel i5-10500K platform with 24 GB RAM.
| Method | SNR = -10 dB | SNR = -5 dB | SNR = 0 dB | SNR = 10 dB |
|---|---|---|---|---|
| FDC | 12.3 | 12.1 | 12.0 | 11.8 |
| FrFT | 108.5 | 107.9 | 106.2 | 105.1 |
| WHT | 15.8 | 15.6 | 15.5 | 15.3 |
| MSDC | 35.2 | 34.8 | 34.5 | 34.1 |
While MSDC is slower than FDC and WHT, its execution time (~35 ms) is still well within the 500 ms real-time requirement for China UAV surveillance. The additional computational cost is justified by the substantial improvement in detection performance, especially under challenging conditions. FrFT, on the other hand, is too slow for real-time deployment.
Comprehensive Radar Chart
To holistically evaluate the methods, we normalize five key metrics: detection probability, precision, recall, false alarm suppression (1 – \( P_{fa} \)), and computational efficiency (1/time). Table IV provides the normalized scores (0 to 1) at SNR = -5 dB.
| Metric | FDC | FrFT | WHT | MSDC |
|---|---|---|---|---|
| Detection Probability | 0.659 | 0.396 | 0.279 | 1.000 |
| Precision | 0.829 | 0.541 | 0.411 | 1.000 |
| Recall | 0.601 | 0.355 | 0.231 | 1.000 |
| False Alarm Suppression | 0.950 | 0.816 | 0.776 | 0.991 |
| Efficiency | 1.000 | 0.113 | 0.774 | 0.348 |
MSDC dominates in detection probability, precision, recall, and false alarm suppression, while being slightly less efficient than FDC and WHT. However, the overall balance is strongly in favor of MSDC for China UAV detection in realistic noisy environments.
Hardware Deployment Feasibility
Our algorithm is designed with low computational complexity \( O(N \log N) \), making it suitable for software-defined radio (SDR) platforms. We analyzed resource usage on an AD9361 + Zynq-7000 FPGA platform: preprocessing and pulse compression modules can be implemented as parallel IP cores, requiring about 12,000 LUTs, 8,000 Flip-Flops, and 512 KB memory. The total single-sample latency is projected to be under 500 ms after optimization, meeting the real-time requirement for monitoring multiple China UAVs simultaneously (up to 3 units).
Conclusion
In this work, we have proposed a robust detection method for Drone-ID signals from DJI China UAVs based on pulse compression and multi-scale dual correlation. The three-stage architecture—preprocessing, dual correlation enhancement, and adaptive feature verification—effectively addresses the challenges of low SNR, frequency offsets, and false peaks. Extensive experiments demonstrate that our method achieves a detection probability of 91.2% at SNR = -5 dB with a false alarm rate below 1%, outperforming traditional approaches by a wide margin. The algorithm is computationally efficient and suitable for real-time deployment on SDR platforms, providing a practical solution for China UAV security and airspace management. Future work will focus on extending the method to handle high-speed China UAV maneuvers and more complex interference scenarios, further improving its adaptability and reliability.