Drone Road Systems in Transport Infrastructure

• Drone Road Systems (DRS) are road-integrated UAV networks that enhance transport infrastructure through coordinated aerial and ground operations.
• DRS research combines sensor technologies, geo-localization, and digital reconstruction methods to support efficient road maintenance and detailed traffic observation.
• DRS architectures integrate adaptive routing, emergency assessment, and decentralized guidance to improve safety and operational efficiency in modern transport networks.

Drone Road Systems (DRS) designate a family of road-coupled unmanned-aerial systems in which drones become operational elements of transport infrastructure rather than merely ad hoc flying sensors. Across the literature, the term covers autonomous road-maintenance platforms in which UAVs coordinate with ground robots, drone-based systems for capturing naturalistic road-user trajectories, weather-robust geo-localization stacks that fuse road maps with satellite imagery, aerial relay infrastructures for vehicular communications and optical backhaul, truck-drone logistics architectures, and explicitly modeled aerial road networks with decentralized guidance rules (Katsamenis et al., 2022, Bock et al., 2019, Janji et al., 2022, Qu et al., 28 Sep 2025). The literature also uses the acronym DRS for “Drone Relay Station” in communications papers, which broadens the term beyond a single canonical architecture.

1. Conceptual scope and terminological variation

The literature does not present a single universally fixed definition of DRS. Instead, it presents several technically related meanings unified by one premise: drones are embedded into road-centered cyber-physical systems. In some works, the road itself is the maintained asset; in others, it is the observed traffic scene, the communication corridor, the logistics graph, or a structured aerial roadway.

DRS sense in the literature	Role of drones	Representative papers
Road-infrastructure maintenance	Inspection, mapping, coordination with UGVs	(Katsamenis et al., 2022)
Traffic observation and trajectory capture	Overhead sensing of multimodal road users	(Bock et al., 2019, Breuer et al., 2020)
Geo-localization and scene reconstruction	Cross-view localization, aerial priors, reconstruction support	(Fang et al., 14 May 2026, Zhang et al., 2024)
Relay infrastructure	FSO backhaul or RIS-assisted V2X reflection	(Janji et al., 2022, Janji et al., 24 Mar 2025, Janji, 25 Mar 2025)
Truck-drone logistics and assessment	Launch, recovery, routing, delivery, PDRA	(Daknama et al., 2017, Yang et al., 2024, Gong et al., 24 Oct 2025, Liu et al., 26 Jun 2026)
Structured aerial road networks	Lane-based airspace representation and decentralized guidance	(Qu et al., 28 Sep 2025)

A common misconception is that DRS denotes only a lane-based airborne traffic system. The surveyed literature suggests a broader interpretation: DRS is an umbrella concept for systems in which drones are structurally coupled to roads, road assets, road users, or road-network operations. A second source of ambiguity is acronym overlap. In vehicular communications, DRS explicitly means “Drone Relay Station,” especially when a UAV carries an FSO transceiver or a reconfigurable intelligent surface (Janji et al., 2022, Janji et al., 24 Mar 2025).

2. System architecture and road coupling

A recurring architectural pattern is multi-layer integration among aerial assets, ground assets, road geometry, communications, and a supervisory software layer. HERON is the clearest maintenance-oriented instantiation: UAVs support visual inspection, mapping, monitoring, and coordination in pre-/post-intervention phases; a UGV performs pothole patching, crack sealing, asphalt rejuvenation, replacement of removable urban pavement elements, road marking painting, and automated traffic-cone handling; sensing interfaces and laser scanners supply 3D mapping; AI toolkits coordinate workflows; control software translates high-level plans into low-level robotic behaviors; and an integrated DSS/IMS with AR/VR-enhanced UI aggregates the Common Operational Picture (Katsamenis et al., 2022). The workflow is explicitly organized as inspection → mapping → defect identification → planning → execution → verification.

Road coupling is frequently realized through explicit geometric reference frames. In the intersection-oriented inD pipeline, stabilized drone imagery is mapped to a local ground-plane coordinate system through a projective homography $H$, with $X \propto Hx$ and normalization $X = \frac{Hx}{h_3^\top x}$, so trajectories become metric and comparable across flights (Bock et al., 2019). In openDD, drone trajectories and HD map layers are jointly represented in UTM coordinates, with lane centerlines, lane boundaries, and drivable areas distributed as aligned shapefiles and XML (Breuer et al., 2020). In both cases, the drone is not an isolated observer; it is a measurement instrument tied to a formally represented road topology.

This road-coupled structure extends beyond sensing. The literature repeatedly combines edge/cloud inference, graph-based planning, and communication backbones. HERON uses Vehicle-to-Infrastructure/Everything exchange over 4G/5G, WiMAX, and BLE4 with a robust self-healing architecture (Katsamenis et al., 2022). Geo-localization systems couple precisely geo-aligned road map tiles to satellite tiles, so the road network becomes a weather-invariant prior rather than a passive background (Fang et al., 14 May 2026). The shared pattern is not generic UAV autonomy, but a road-referenced stack in which geometry, task allocation, and decision support are all indexed to transport infrastructure.

3. Sensing, mapping, traffic observation, and digital twins

One major DRS lineage uses drones as overhead sensing platforms for road-state estimation and traffic analytics. The inD dataset exemplifies this approach at urban intersections. It uses a DJI Phantom 4 Pro multicopter with gimbal stabilization, 4K recording at $4096\times2160$ pixels and 25 frames per second, hovering at up to 100 m altitude to cover approximately $80\times40$ m to $140\times70$ m per site. The released dataset contains more than 11,500 road users, including more than 5,000 vulnerable road users, over 10 hours at four unsignalized intersections in Aachen (Bock et al., 2019). Its extraction pipeline favors semantic segmentation over bounding-box detection, trains two U-Net-based networks for small and large objects, performs nearest-neighbor tracking with short-occlusion prediction, and applies Bayesian smoothing with a constant-acceleration model, reducing positioning error to approximately one pixel, about 4 cm at typical altitude.

openDD extends the same overhead-trajectory paradigm to roundabouts at much larger scale. It provides 84,774 accurately tracked trajectories over 62.7 hours from 501 drone flights across seven roundabouts, with 3840×2160 px stabilized footage at 30 fps, UTM-referenced state vectors, and HD maps containing lane centerlines, boundaries, and drivable areas (Breuer et al., 2020). The dataset thereby supports lane-level occupancy analysis, yielding and merging studies, trajectory prediction, and conflict analytics in unsignalized circulatory traffic. Together, inD and openDD establish a DRS design pattern in which drones supply low-occlusion, wide-area, naturalistic trajectory data that roadside sensing often struggles to obtain.

A second observation-oriented branch emphasizes real-time incident detection. DARTS integrates thermal imaging, YOLO-based vehicle detection, Lucas–Kanade tracking, monochrome trajectory-image generation, and the TCD-Net classifier to distinguish normal conditions, recurrent congestion, and incident-induced congestion. The system records 2-minute segments in three concurrent threads staggered every 40 seconds, achieved 99% detection accuracy on a self-collected dataset, and in a field test on Interstate 75 detected and verified a rear-end collision 12 minutes earlier than the local transportation management center (Li et al., 29 Oct 2025). Because it also estimates incident-induced congestion length from GPS spans across consecutive incident segments, it links aerial observation to operational response rather than mere detection.

A third sensing lineage treats drones as complementary views for digital reconstruction. Drone-assisted road Gaussian Splatting introduces cross-view uncertainty into 3D-GS, using ensemble-based ground-view uncertainty projected onto aerial pixels so that aerial images contribute primarily where car-view learning is weak. On held-out road views, it improves over ground-only Scaffold-GS(G), including PSNR gains of 0.68 in the NYC scene and 0.41 in the SF scene, with additional gains under small viewpoint shifts and tilts (Zhang et al., 2024). This places DRS in the emerging domain of road-scene digital twins for autonomous-driving simulation.

4. Localization, relay infrastructure, and vehicular communications

Another branch of DRS research treats roads as geometric priors or communication corridors rather than maintenance targets. GeoFuse frames DRS as a geo-localization stack in which road maps and building footprints are “free geometric priors” fused with satellite imagery to stabilize drone-view localization under rain, fog, snow, illumination shifts, and blur. The system uses X-VLM with a Swin Transformer image encoder and BERT text encoder, shared image encoders for satellite and road map tiles, dual-level token/channel fusion, and class-level cross-view contrastive learning. It reports mean Recall@1 improvements of +3.46% on University-1652 and +23.18% on DenseUAV, indicating that road topology can function as a strong weather-invariant cue for aerial localization (Fang et al., 14 May 2026).

In communications, DRS frequently denotes a UAV relay. For FSO backhaul, a drone relay station is placed at geometrically valid points derived from building vertices and sunny points in high-rise urban areas, with Lee’s visibility-graph algorithm and Dijkstra search producing unobstructed line-of-sight relay chains between a macro base station and a hotspot. A solar-aware placement strategy prioritizes sunlit nodes, and the reported 24-hour Madrid simulation reduces recharge trips from 115 without PV to around 74 with PV, approximately 35% fewer trips (Janji et al., 2022). The paper explicitly omits detailed FSO channel and throughput modeling, so its contribution is geometric placement and sustainability rather than end-to-end optical PHY design.

RIS-assisted vehicular communications extend the same relay logic. One study mounts a reconfigurable intelligent surface on a UAV and optimizes DRS trajectory toward the geometric midpoint between communicating vehicles, using Q-learning to control RIS orientation; the reported result is consistent path-loss reduction and increased rate, especially as direct links degrade with distance or blockage (Janji et al., 24 Mar 2025). A related study derives an analytical yaw control that suppresses interference by nulling the interferer’s RIS array factor; its simulations show a modest average throughput improvement of about 0.5–0.6% over trajectory optimization without orientation control (Janji, 25 Mar 2025). In both cases, the road becomes a dynamic RF corridor, and the DRS is an aerial propagation-control device rather than a sensor platform.

5. Routing, logistics, and emergency-assessment systems

A large part of the DRS literature models drones as tightly coupled to road-network routing problems. In “Vehicle Routing with Drones,” trucks act as mobile launch, landing, and charging platforms. Drones can ride on trucks or fly, but while flying they can carry only one package at a time and must return to a truck to charge after each delivery. The paper’s nested local-search heuristic outperforms a natural Greedy baseline and reports substantial savings relative to truck-only delivery, including a 21.3% reduction in average delivery time for 1 truck and 2 drones with 200 packages (Daknama et al., 2017). Here the “road system” is embodied by the truck route itself: the truck becomes a moving aerial-service corridor.

The road-network version of this idea appears in MA-FSTSP, which embeds multiple trucks and multiple drones in a strongly connected directed road graph with one-way arcs. Trucks must move on the road network, drones may fly directly between vertices subject to endurance, and launch/recovery must synchronize with truck presence. The proposed mixed-integer model and three-phase heuristic outperform column generation and variable neighborhood search baselines, and the approach scales to more than 300 customers within a 5-minute time limit (Yang et al., 2024). The road network is no longer an abstract metric background; it explicitly constrains truck routing and the feasible geometry of launch and rendezvous.

Locker-based LTDRP-PDNF pushes the same coupling toward automated service infrastructure. Smart lockers serve as temporary parcel storage facilities and as automated drone docking and service nodes supporting takeoff, landing, parcel handover, and battery replacement. The model integrates deliveries, return pickups, battery-constrained and load-dependent flights, and no-fly-zone detours, with NFZs represented as regular hexagons and drone leg energy given by $$\Delta\beta_{dr,k}^{t} = \alpha(w_{dr,s}+w_{dr,k}^{t})\, r^{dr}(c_{dr,k}^{t},a_{dr,k}^{t})$$ (Liu et al., 26 Jun 2026). The two-stage DRL heuristic produces the best reported objectives for $N=50$ and $N=100$ instances while retaining short computation times of 3.95 s and 14.79 s, respectively.

Emergency and assessment applications use similar road-graph abstractions. The unified model for post-disaster road assessment transforms a link-based road network $G=(N,A)$ into $X \propto Hx$0, where artificial nodes represent assessable road segments with information value. A single transformer encoder–decoder is trained across eight PDRA variants, reaches real-time inference of 1–10 seconds on networks up to 1,000 nodes, reduces training time and parameters by a factor of eight relative to separate models, and outperforms single-task DRL by 6–14% and traditional optimization by 24–82% in collected information value (Gong et al., 24 Oct 2025). In the dual-task urban monitoring setting, drones carrying parcels also refresh traffic information on road segments whose uncertainty grows with time since last observation. On the Barcelona network, the decentralized “meet-and-merge” method achieves information gain of about 789,091, coverage of about 80.67%, and average AoI of about 46.46%, while using about 609 MILP calls instead of about 1,655 for the centralized baseline (Zhu et al., 2 Apr 2026). These systems suggest that DRS can treat the road network itself as an information field.

6. Structured aerial roads, guidance, evaluation, and open issues

The most literal interpretation of DRS is a formally specified aerial road network. The short-term guidance literature defines a DRS as a set of oriented, uni-directional roads in 3D Cartesian space, each with a center lane and multiple parallel lanes arranged in a hexagonal lattice in the normal plane of a $X \propto Hx$1 center curve. Roads are linked by on-ramps, off-ramps, and connecting ramps, and the entire system is machine-readable through an XML markup language with Curve, ChainedCurve, Lane, Road, Ramp, and DroneRoadSystem elements (Qu et al., 28 Sep 2025). This is not merely an analogy to roads; it is an explicit airborne roadway formalism.

Guidance within that formalism is decentralized. The Short-Term Decentralized Greedy algorithm uses only periodically transmitted beacons carrying position, velocity, road identifier, lane identifier, and lane parameter. It then selects among the current lane and immediate neighboring lanes, minimizing a cost composed of speed deviation, target-lane position cost, and collision cost. Simulation varied beacon rates of 5, 20, and 50 Hz and transmit powers of 2 mW and 20 mW. The maximum observed collision rate was 0.567 at 50 Hz and 20 mW, while 2 mW with an intermediate 10 Hz setting was identified as a balanced operating point (Qu et al., 28 Sep 2025). The paper also reports an “inner-lane vanishing” phenomenon under high congestion, where inner lanes become underutilized because their higher neighbor density inflates collision cost.

Across the broader DRS literature, several limitations recur. HERON remains semi-automated, with humans in the loop and some autonomy details still unspecified; the paper does not specify detailed UAV sensor payloads, exact multi-agent task-allocation algorithms, or quantitative field performance metrics (Katsamenis et al., 2022). inD is restricted to daylight, low-wind, unsignalized intersections in one city, and does not measure object height directly (Bock et al., 2019). The FSO relay work deliberately omits channel, SNR, BER, and throughput modeling, focusing instead on geometry and trip-count sustainability (Janji et al., 2022). The literature therefore suggests that DRS is still better understood as a heterogeneous research program than as a fully standardized infrastructure class.

A further implication is that DRS research balances two partially conflicting ambitions. One is strict structure: lane models, road graphs, digital twins, relay corridors, and formal optimization. The other is adaptive opportunism: drones as mobile sensors, movable relays, and dynamic launch platforms. The field’s strongest results typically occur when both are combined—when road geometry is explicit, but aerial decisions remain adaptive to defects, traffic, weather, obstruction, or demand.