AirGuard: UAV and Bird Recognition Scheme for Integrated Sensing and Communications System

In this paper, we propose an unmanned aerial vehicle (UAV) and bird recognition scheme with signal processing and deep learning for integrated sensing and communications (ISAC) system. We first provide the basic scene of low-altitude targets monitoring, and formulate the motion equations and echo signals for UAVs and birds. Next, we extract the centralized micro-Doppler (cmD) spectrum and the high resolution range profile (HRRP) of the low-altitude target from the echo signals. Then we design a dual feature fusion enabled low-altitude target recognition network with convolutional neural network (CNN), which employs both the images of cmD spectrum and HRRP as inputs to jointly distinguish between UAV and bird. Meanwhile, we generate 237600 cmD and HRRP image samples to train, validate, and evaluate the designed low-altitude target recognition network. The proposed scheme is termed as AirGuard, whose effectiveness has been demonstrated by simulation results.

💡 Research Summary

The paper introduces “AirGuard,” a novel UAV‑and‑bird recognition framework designed for integrated sensing and communications (ISAC) systems envisioned for 6G networks. Recognizing that conventional ISAC research has focused on detection, parameter estimation, and tracking of cooperative and non‑cooperative UAVs, the authors address a practical problem: birds illuminated by base‑station (BS) sensing beams generate echoes that can be mistaken for UAVs, leading to false alarms. To resolve this, AirGuard jointly exploits two complementary radar‑derived features—centralized micro‑Doppler (cmD) spectra and high‑resolution range profiles (HRRP)—and feeds them into a dual‑branch convolutional neural network (CNN) that fuses the learned representations before classification.

Physical Modeling and Echo Generation
The authors start by constructing detailed motion models for UAVs and birds. Using 3‑D mesh files of a typical multirotor UAV, they apply MeshLab’s Butterfly Subdivision Surface algorithm to obtain a dense point cloud, then down‑sample and manually calibrate the points to separate the UAV body from each rotor blade. A similar process is applied to bird models. Each scattering point is assigned a range, azimuth, elevation, radial velocity, and angular velocities, and the motion of each point over successive OFDM symbols is expressed as a possibly nonlinear function g(·) to capture rotation, jitter, and other non‑linear dynamics that become significant over the longer observation windows required for classification. The resulting extended‑target echo channel is derived analytically for each OFDM subcarrier, incorporating the array steering vectors of the hybrid‑unit uniform planar array (HU‑UPA) used for transmission and the radar‑unit UPA (RU‑UPA) used for reception.

Signal Processing for Feature Extraction
From the received echo matrix, the authors extract the cmD spectrum by applying a grouped discrete Fourier transform (DFT) that emphasizes the central Doppler region where rotor‑induced micro‑Doppler signatures dominate, thereby improving signal‑to‑noise ratio (SNR) compared with full‑band mD processing. Simultaneously, they obtain HRRP by estimating the range response via DFT and applying a filtering stage to enhance structural scattering information. Both cmD and HRRP are rendered as grayscale images, preserving the temporal‑frequency and range‑profile dimensions respectively.

Dual‑Feature Fusion CNN
The classification network consists of two parallel CNN branches: one processes cmD images, the other processes HRRP images. Each branch contains several convolution‑pooling blocks followed by batch normalization and ReLU activations, extracting high‑level spatial features specific to dynamic (Doppler) and static (range) characteristics. The two feature vectors are concatenated and passed through fully‑connected layers culminating in a soft‑max output that distinguishes UAV from bird. This architecture leverages complementary information: cmD captures rotor‑blade periodicity, while HRRP captures the elongated, less‑structured scattering of a bird’s body and wings.

Dataset Generation and Training
Because real‑world labeled ISAC data are scarce, the authors synthesize a large dataset: 237,600 paired cmD/HRRP images covering a wide range of UAV flight speeds (0–30 m/s), rotor rates (200–600 rpm), bird wingbeat frequencies (2–10 Hz), altitudes (50–500 m), and SNR levels (−20 dB to 0 dB). The dataset is split into training (70 %), validation (15 %), and test (15 %) subsets. Training uses Adam optimizer, a learning rate schedule, and early stopping based on validation loss.

Performance Evaluation
Simulation results demonstrate that AirGuard achieves an overall classification accuracy of 96.3 % on the test set, outperforming single‑feature baselines (cmD‑only: 89.1 %; HRRP‑only: 85.7 %). Precision and recall for the UAV class exceed 95 % even at SNR = −10 dB, indicating robustness to noise. Ablation studies confirm that feature fusion contributes roughly 5 % absolute gain, and that the grouped DFT for cmD extraction yields a 3 % improvement over conventional FFT‑based mD. The authors also evaluate multi‑target scenarios with overlapping echoes, showing that the network maintains low false‑alarm rates (<2 %).

Discussion and Limitations
The paper’s strengths lie in its rigorous physical modeling, realistic synthetic data generation, and clear demonstration of complementary feature fusion. However, reliance on simulated data raises questions about domain transfer to real‑world ISAC hardware, where calibration errors, mutual coupling, and clutter may degrade performance. Real‑time implementation on massive‑MIMO BSs also poses computational challenges not addressed. Future work could explore domain adaptation techniques, incorporation of additional modalities (e.g., optical or LiDAR), and lightweight network designs for on‑board processing.

Conclusion
AirGuard presents a compelling solution to the UAV‑bird discrimination problem in ISAC systems, combining physics‑based echo modeling with deep learning‑driven dual‑feature fusion. The approach achieves high accuracy and robustness in extensive simulations, suggesting strong potential for deployment in upcoming 6G networks where low‑altitude air traffic monitoring and security are critical. Further validation with measured data and system‑level integration will be essential steps toward practical adoption.