Relaying Signal When Monitoring Traffic: Double Use of Aerial Vehicles Towards Intelligent Low-Altitude Networking

In intelligent low-altitude networks, integrating monitoring tasks into communication unmanned aerial vehicles (UAVs) can consume resources and increase handoff latency for communication links. To address this challenge, we propose a strategy that enables a “double use” of UAVs, unifying the monitoring and relay handoff functions into a single, efficient process. Our scheme, guided by an integrated sensing and communication framework, coordinates these multi-role UAVs through a proactive handoff network that fuses multi-view sensory data from aerial and ground vehicles. A lightweight vehicle inspection module and a two-stage training procedure are developed to ensure monitoring accuracy and collaborative efficiency. Simulation results demonstrate the effectiveness of this integrated approach: it reduces communication outage probability by nearly 10% at a 200 Mbps requirement without compromising monitoring performance and maintains high resilience (86% achievable rate) even in the absence of multiple UAVs, outperforming traditional ground-based handoff schemes. Our code is available at the https://github.com/Jiahui-L/UAP.

💡 Research Summary

The paper tackles the growing need for unmanned aerial vehicles (UAVs) to serve dual functions in low‑altitude networks: acting as communication relays and performing traffic monitoring. Conventional approaches treat these tasks separately, leading to excessive resource consumption, high handoff latency, and the need for costly beam‑sweeping or channel estimation before each handoff decision. To overcome these limitations, the authors propose the Unified Aerial Perception Network (UAP‑Net), an integrated sensing‑and‑communication (ISAC) framework that merges monitoring and handoff processes into a single, proactive operation.

System Model
The considered scenario comprises K ground base stations (RSUs) equipped with large uniform linear arrays (ULAs), M UAV relays each with smaller ULAs, and V single‑antenna ground users. Each UAV carries five RGB cameras (front, back, left, right, downward) while each vehicle is equipped with a LiDAR sensor. Communication follows a decode‑and‑forward (DF) protocol; users can receive data either directly from an RSU (direct link, DL) or via a UAV relay (indirect link, IL). The achievable rate for each link is expressed analytically, and the optimal link selection problem is cast as a maximization of the sum‑rate across all users.

UAP‑Net Architecture
UAP‑Net consists of three main modules:

1. UAV Feature Extraction (UFE): Raw RGB images are Z‑score normalized per channel, down‑sampled to 10 % of their original resolution, and fed into a lightweight ResNet‑18 backbone. The five view‑specific feature vectors are concatenated and passed through an additional convolutional block, yielding a compact visual representation (x_{RGB}^{m}) for UAV m.

2. Vehicle Feature Extraction (VFE): LiDAR point clouds are voxelized into a 3‑D binary grid; voxels containing an RSU are labeled with a unique identifier to embed semantic information. A custom 3‑D CNN processes this grid and outputs a vehicle‑centric representation (x_{LiDAR}^{v}).

3. Adaptive Cross‑Agent Fusion (ACAF): To fuse heterogeneous modalities, a cross‑attention mechanism is employed. A learnable cooperation token (x_{Coop}) serves as the query, while each agent’s feature acts as key and value. Type embeddings (LiDAR, RGB, Coop) and sinusoidal positional embeddings encode modality and view order, allowing the fusion to adapt to a varying number of UAV views. The attention output (h_{Coop}) is fed to a fully‑connected layer that predicts the best handoff decision (\hat{\kappa}).

Two networks are trained jointly but in a staged manner:

• Stage‑I (handoff training) optimizes the handoff network (G(\cdot)) using a cross‑entropy loss over the set of candidate links. This stage learns to map the fused multi‑agent features to the optimal link without explicit channel estimation.

• Stage‑II (monitoring training) freezes the UAV feature extractor and trains a lightweight Traffic Inspection Head (TIH) that regresses lane‑wise vehicle counts from (x_{RGB}^{m}). The loss is mean‑squared error, ensuring that the monitoring task does not degrade the handoff performance (negative transfer).

Distributed Execution
After centralized training, the UFE runs on each UAV, continuously extracting visual features that are simultaneously used by the TIH for monitoring and broadcast to nearby vehicles. Vehicles run the VFE locally, fuse their LiDAR representation with the received UAV features via ACAF, and execute the proactive handoff decision before link quality deteriorates. This distributed scheme eliminates the need for exhaustive beam training and reduces latency dramatically.

Experimental Evaluation
The authors evaluate UAP‑Net on the M3SC low‑altitude urban dataset, which includes dense building layouts and frequent non‑line‑of‑sight conditions. The testbed consists of 4 RSUs (128‑element ULAs), 6 UAVs (32‑element ULAs), and 20 possible RSU‑UAV‑user links. Sensors operate at 20 Hz with a 10 % artificial data‑loss to emulate sensor failures; the carrier frequency is 28 GHz.

Key results:

• At a target throughput of 200 Mbps, the communication outage probability drops by roughly 10 % compared with traditional handoff schemes that rely on separate sensing and channel estimation.

• When only a single UAV is available (i.e., the multi‑UAV redundancy is removed), the system still achieves an average achievable rate of 86 % of the multi‑UAV baseline, demonstrating strong resilience.

• Monitoring accuracy (lane‑wise vehicle count) remains on par with or slightly better than dedicated single‑task models, confirming that the shared feature backbone does not compromise perception quality.

Implementation details include training on an NVIDIA RTX 3090 GPU, gradient accumulation with a step size of 8 to manage memory, and a total parameter count below 2 M to suit embedded hardware.

Contributions and Impact
The paper makes several notable contributions:

1. Introduces a novel ISAC‑driven framework that unifies handoff prediction and traffic monitoring for low‑altitude UAV networks.

2. Designs a modular multi‑modal fusion architecture (ACAF) capable of handling dynamic numbers of UAV views and heterogeneous data types.

3. Proposes a two‑stage training strategy that mitigates negative transfer between communication and perception tasks.

4. Demonstrates, via realistic simulations, that proactive, sensor‑driven handoff can substantially reduce outage probability while maintaining high monitoring fidelity.

Overall, the work provides a practical pathway toward smarter, more efficient aerial‑ground networks in future smart‑city deployments, where UAVs can simultaneously ensure reliable connectivity and deliver real‑time situational awareness without incurring prohibitive computational or latency costs.