The rapid expansion of the low-altitude economy, driven by the deregulation of airspace, has positioned unmanned aerial vehicles (UAVs) as a cornerstone of modern logistics, surveillance, and urban air mobility. However, the proliferation of ‘drone technology’ brings forth significant safety challenges, particularly the risk of collisions between UAVs and manned aircraft, as well as unauthorized “black flights” over sensitive infrastructure. To ensure the safe and efficient operation of low-altitude airspace, reliable and robust detection systems are paramount. Traditional single-sensor detection methods, such as radar, acoustic, or visual systems, each suffer from inherent limitations. Radar struggles with low-flying, small UAVs; visual systems are vulnerable to weather and occlusion; and acoustic sensors have limited range. This research addresses these challenges by proposing a novel, high-precision detection algorithm that leverages the complementary strengths of dual-modal data: radio frequency (RF) signals and visual imagery. Our framework employs a cross-attention mechanism for feature fusion, significantly enhancing detection accuracy and robustness over single-modality approaches.
Our proposed detection framework operates by receiving both the RF signals emitted by a drone’s communication system and the visual data from a camera. The core innovation lies in the cross-attention fusion network, which intelligently combines features extracted from each modality. The process begins with separate preprocessing and feature extraction pipelines for RF and image data. The RF signal undergoes a series of transformations to highlight its unique characteristics, while the image data is processed by a state-of-the-art convolutional neural network. The resulting feature maps are then fed into the cross-attention module, which dynamically weights the importance of features from one modality based on information from the other, creating a comprehensive and context-aware representation of the environment. This fused representation is then used for the final classification decision, determining whether a drone is present.
1. RF Signal Processing and Feature Extraction
The accurate analysis of RF signals is critical for ‘drone technology’ detection, as these signals are less susceptible to environmental conditions like fog or lighting. Our methodology for processing low-altitude UAV RF signals involves a multi-stage pipeline designed to denoise, compress, and extract salient features.
1.1. Signal Model and Denoising
The received RF signal, \( r(t) \), is modeled as a superposition of the transmitted signal \( s(t) \), additive white Gaussian noise \( n_{AWGN}(t) \), and co-channel interference \( n_{inter}(t) \). This is represented by the equation:
$$ r(t) = [s(t)*h(t)] + n_{AWGN}(t) + n_{inter}(t) $$
where \( h(t) \) is the channel impulse response. To isolate the drone’s signal from background noise, we employ Ensemble Empirical Mode Decomposition (EEMD). EEMD is superior to traditional methods like Fourier or wavelet transforms for analyzing non-linear, non-stationary signals, as it adaptively decomposes the signal into intrinsic mode functions (IMFs) without requiring pre-defined basis functions. The relationship between the added white noise amplitude \( A \) and the number of integration trials \( N \) is given by:
$$ \varepsilon = \frac{A}{\sqrt{N}} $$
where \( \varepsilon \) is the standard deviation of the added noise. This process effectively separates high-frequency noise from the signal’s core components.
1.2. Compressed Sensing and Sparsity
UAV communication signals exhibit a high degree of redundancy and sparsity in a transformed domain, a property exploited by compressed sensing theory. We validate the low-rank nature of low-altitude UAV RF signals, showing that energy is concentrated in a small fraction of the data. If a signal \( X \in \mathbb{R}^N \) is K-sparse in a transform domain \( \Psi \), it can be represented as:
$$ X = \Psi S $$
where \( S \) has at most K non-zero entries. The core idea of compressed sensing is to capture a limited number of measurements \( Y \in \mathbb{R}^M \) (where \( M \ll N \)) that contain enough information to reconstruct the original signal:
$$ Y = \Phi X $$
Here, \( \Phi \) is an \( M \times N \) measurement matrix that must be incoherent with the sparsifying transform \( \Psi \). In our implementation, we project the RF signal onto the wavelet domain and use a partial Fourier measurement matrix, achieving a compression ratio of approximately 10%. This drastically reduces the computational load for subsequent processing stages. The complexity of our proposed signal processing pipeline is compared with the traditional approach in the table below, illustrating a 39% reduction in algorithmic complexity.
| Processing Stage | Traditional Complexity | Proposed Complexity (with CS) |
|---|---|---|
| EEMD | \( O(N \cdot M \cdot K) \) | \( O(N \cdot M \cdot K) \) |
| STFT | \( O(N \cdot W \cdot \log W) \) | \( O(N \cdot W \cdot \log W) \) |
| Compressed Sensing | N/A | \( O(N \log N) \) |
| Total (Example: N=10^6) | \( 2 \times 10^8 \) FLOPs | \( 1.22 \times 10^8 \) FLOPs |
1.3. Time-Frequency Analysis
To visualize the temporal evolution of the signal’s frequency components, we perform a Short-Time Fourier Transform (STFT). The STFT divides the signal into short, overlapping time windows and applies a Fourier transform to each. The mathematical definition is:
$$ STFT_x(t, f) = \int_{-\infty}^{\infty} x(\tau) w(\tau – t) \cdot e^{-j2\pi f\tau} d\tau $$
where \( x(\tau) \) is the input signal and \( w(\tau – t) \) is the windowing function. The output of the STFT is used to generate two distinct 2D representations: the spectrogram, which displays the power spectral density over time, and the persistence spectrum, which highlights the most frequently occurring frequency components. These images serve as the input to our deep learning classifier.
1.4. Feature Extraction with a Residual Neural Network
For the classification of the RF signal images, we utilize a Residual Network (ResNet) architecture. ResNet is renowned for its ability to train very deep networks by introducing skip connections. These connections allow the network to learn an identity mapping, effectively mitigating the vanishing gradient problem. The fundamental building block of a ResNet, such as ResNet-50, consists of a series of convolutional layers with a shortcut connection that bypasses the convolutional operations. The output of a residual block can be expressed as:
$$ y = F(x, \{W_i\}) + x $$
where \( x \) is the input, \( F(x, \{W_i\}) \) represents the residual mapping to be learned, and \( y \) is the output. Our experiments, detailed in the results section, show that ResNet-50 achieves high classification accuracy for both spectrogram and persistence spectrum images across various classification tasks (2-class, 4-class, and 10-class).
2. Image Data Feature Extraction and Object Detection
Visual data provides rich contextual and spatial information that is complementary to RF signals. We employ a two-stage approach: an EfficientNet backbone for feature extraction and an EfficientDet head for object detection and localization.
2.1. EfficientNet for Feature Extraction
EfficientNet is a family of convolutional neural networks that achieve state-of-the-art performance with high efficiency. Its core innovation is a compound scaling method that uniformly scales the network’s depth, width, and input resolution. This scaling is governed by the formula:
$$ depth: d = \alpha^\phi, \quad width: w = \beta^\phi, \quad resolution: r = \gamma^\phi $$
Here, \( \phi \) is a user-specified scaling coefficient, and \( \alpha, \beta, \gamma \) are constants determined through a neural architecture search, with the constraint \( \alpha \cdot \beta^2 \cdot \gamma^2 \approx 2 \). This ensures that the computational cost (FLOPs) increases in a controlled manner while maximizing accuracy. We modify the EfficientNet-b1 architecture by replacing its standard Squeeze-and-Excitation (SE) attention module with a Coordinate Attention (CA) module. CA decomposes channel attention into two 1D feature encoding processes, which helps the network to capture long-range spatial dependencies, a critical factor for detecting small objects like drones.
2.2. EfficientDet for Object Detection
Building upon the EfficientNet backbone, the EfficientDet detection framework employs a weighted Bi-directional Feature Pyramid Network (Bi-FPN) for multi-scale feature fusion. The Bi-FPN introduces learnable weights for the feature fusion process, allowing the network to dynamically prioritize the most informative scales. The fusion of features is mathematically represented as:
$$ \tilde{U} = \sum_{i=1}^{n} \frac{w_i}{\epsilon + \sum_{j} w_j} U_i $$
where \( \tilde{U} \) is the fused feature, \( U_i \) are the input features, \( w_i \) are their corresponding learnable weights, and \( \epsilon \) is a small constant to prevent division by zero. For bounding box regression, we use Wise-IoU (WIoU) as the loss function, which is particularly effective for small objects. The WIoU loss includes a dynamic focusing mechanism to prevent detrimental gradients from low-quality anchors. Its formulation is:
$$ \mathcal{L}_{WIoU} = \mathcal{R}_{WIoU} \cdot \mathcal{L}_{IoU} $$
where \( \mathcal{L}_{IoU} \) is the standard IoU loss and \( \mathcal{R}_{WIoU} \) is a penalty term that amplifies gradients for high-quality anchors and suppresses them for low-quality ones. This ensures the model focuses on learning from the most relevant examples.
3. Cross-Attention Multimodal Fusion Network
The core of our contribution is the cross-attention mechanism that intelligently fuses the features from the RF signal and vision modalities. This approach allows the model to learn which features from one modality are most relevant to the other, creating a robust and context-aware representation. In our 8-head cross-attention fusion network, the RF feature sequence \( F_r \in \mathbb{R}^{D_r} \) serves as the Query, while the image feature sequence \( F_i \in \mathbb{R}^{D_i} \) serves as both the Key and the Value. A linear projection is first applied to project the features into a common dimensionality \( D_h \):
$$ Q = W_q F_r + b_q, \quad K = W_k F_i + b_k, \quad V = W_v F_i + b_v $$
where \( W_q, W_k, W_v \) are learnable weight matrices. The attention weights \( A \) are then computed using the scaled dot-product attention mechanism:
$$ A = \text{softmax}\left(\frac{QK^T}{\sqrt{D_h}}\right) $$
The final fused feature \( F_{fusion} \) is obtained by combining the original RF query with the resulting attention-weighted value, followed by layer normalization:
$$ F_{fusion} = \text{LayerNorm}(F_r + AV) $$
This process is performed in parallel across multiple “heads,” each learning different aspects of the cross-modal relationship. The outputs from all heads are concatenated and passed through a final linear projection to produce the fused feature vector, which is then used for final classification.
4. Experiments and Results
We conducted extensive experiments using public datasets, DroneRF for RF signals and a Drone Dataset for images, to validate the performance of our individual modules and the overall fused system.
4.1. RF Signal Classification Results
The ResNet-50 model was trained and tested on 2-class, 4-class, and 10-class classification tasks using both spectrogram and persistence spectrum images. The following table summarizes the top-level performance achieved when using multiple signal frames.
| Image Type | Accuracy (%) | F1-Score (%) | ||||
|---|---|---|---|---|---|---|
| 2-Class | 4-Class | 10-Class | 2-Class | 4-Class | 10-Class | |
| Spectrogram | 100 | 97.78 | 78.89 | 100 | 87.93 | 97.59 |
| Persistence | 100 | 100 | 87.78 | 96.77 | 100 | 97.59 |
These results demonstrate that our RF-based classifier is highly effective, achieving near-perfect accuracy for simpler tasks and strong performance for more complex, multi-class scenarios.
4.2. Image Classification and Detection Results
The EfficientDet-b1 model, using our modified EfficientNet backbone, was trained for object detection on the Drone Dataset. The model’s performance on the test set is summarized in the table below.
| Dataset Split | AP | AP50 |
|---|---|---|
| Test Set | 37.4 | 48.1 |
| Validation Set | 36.9 | – |
Furthermore, the model achieved a classification accuracy and F1-Score of 99.8%, demonstrating that the visual modality alone is highly reliable in controlled conditions. The confusion matrix from the classification task verified the model’s effectiveness.
4.3. Multimodal Fusion and Robustness
The ultimate test of our algorithm is the performance of the cross-attention fusion network. When tested on the aligned multimodal dataset, the fused model achieved a classification accuracy of 100% and an F1-Score of 100%. To test robustness, we evaluated the algorithm on unseen, more challenging data. The RF feature extraction network, when tested on the DroneRFa dataset, achieved an accuracy of 97.5%, a recall of 100%, and an F1-Score of 98.7%. Similarly, the overall fused framework, when tested on a new visual dataset, achieved an accuracy of approximately 92.3% and an F1-Score of 93.4%. This slight drop in performance is expected due to domain shift, but the system’s ability to maintain high accuracy in novel environments confirms its robustness. The following table summarizes the robustness test results.
| Module | Test Dataset | Accuracy (%) | Recall (%) | F1-Score (%) |
|---|---|---|---|---|
| RF Classifier (ResNet-50) | DroneRFa | 97.5 | 100 | 98.7 |
| Fused Framework | Unseen Visual Dataset | 92.3 | – | 93.4 |
5. Complexity Evaluation
Beyond accuracy, the computational efficiency of a real-time detection system for ‘drone technology’ is critical. The total FLOPs of our proposed network is calculated by summing the contributions of the feature extractors and the fusion network. The RF feature extraction network (ResNet-50) has approximately 4.12 GFLOPs, and the image feature extraction network (EfficientDet-b1) has about 5.1 GFLOPs. The computational cost of the 8-head cross-attention fusion network is significantly lower and can be approximated by:
$$ O_{Total}^{Fusion} = O_{FC} + O_{Multi} + O_{Attention} + O_{Output} $$
where the individual components are functions of the input and output dimensions (\( d_{model} \)), the number of heads (\( h \)), and key/query dimensions (\( d_k \)). The total FLOPs for the fusion network is calculated to be \( 6.8 \times 10^7 \), which is negligible compared to the feature extractors. Therefore, the overall network complexity is dominated by the feature extractors, totaling approximately 9.2 GFLOPs. This complexity is suitable for deployment on edge computing platforms with modern GPUs.
6. Conclusion
In this work, we presented a high-precision detection algorithm for low-altitude unmanned aerial vehicles, leveraging a novel cross-attention multimodal fusion framework that robustly combines RF signal and visual information. The core strength of our approach lies in its ability to overcome the inherent weaknesses of single-sensor systems. The RF signal provides resilience against lighting and weather, while the visual modality offers spatial and contextual verification. Our results conclusively show that the fused model achieves a perfect 100% classification accuracy on our primary test set, significantly outperforming individual modality classifiers. Furthermore, the robustness tests on new datasets demonstrate the algorithm’s generalization capability and low false-positive rate, making it a reliable solution for practical deployment. The computational complexity analysis confirms that the system can be implemented for real-time operations. This research provides a strong technical foundation for the future of ‘drone technology’ integration into low-altitude airspace management. Future work will focus on extending the fusion framework to not only detect but also localize and track UAVs by using the RF signal to estimate the angle of arrival, thereby directly assisting the camera system in its tracking function.