Comparative Analysis of Map-Matching Algorithms for Autonomous Vehicles Under Varying GPS Errors and Network Densities

Featured Application

The proposed infrastructure-side map-matching framework can be directly applied to cloud-assisted traffic-signal information services for low-cost autonomous vehicles. By accurately determining each vehicle’s approach link using only SBAS-aided GNSS and SD maps, the system enables reliable traffic-signal information delivery in V2N architectures without requiring HD maps or expensive on-board sensors. This capability supports scalable deployment of cooperative intersection-safety services for shared shuttles, delivery robots, and other cost-constrained autonomous mobility platforms.

Abstract

Reliable traffic-signal information delivery is critical for safe navigation through signalized intersections, particularly for low-cost autonomous vehicles that rely on Vehicle-to-Network (V2N) communication rather than on-board HD maps or expensive perception sensors. Ensuring this selective delivery requires accurate infrastructure-side map-matching, which becomes challenging when vehicles operate with only Standard Definition (SD) maps and noisy GNSS measurements. This study comparatively evaluates five infrastructure-side map-matching algorithms under varying GNSS errors and road-network densities using real trajectories from Jeju Island with controlled Gaussian perturbations. The framework includes geometric matching, Extended Kalman Filtering (EKF), route-constrained filtering, grid-based spatial indexing, and a hybrid route–EKF fallback mechanism, executed in real time on a cloud-hosted Kafka pipeline. The hybrid route–EKF algorithm exhibited consistently high and stable link-matching accuracy (0.99308–0.96546 across GPS error groups; 0.9887–0.9777 across density groups) together with strong signal-matching accuracy (0.99394–0.96950; 0.9865–0.9790). Route-constrained and Kalman-based approaches also performed well, while heading-based matching showed clear limitations. These results indicate that infrastructure-side map-matching provides a scalable foundation for cloud-assisted traffic-signal information services and supports the feasibility of delivering reliable traffic-signal information to low-cost autonomous platforms.

1. Introduction

Although precise localization remains an essential element of autonomous driving, the achievable accuracy varies substantially across market segments. Premium robotaxi fleets routinely achieve sub-decimeter accuracy by combining high-definition (HD) maps, multi-sensor fusion, and centimeter-level GNSS corrections such as Real-Time Kinematic (RTK) or Precise Point Positioning (PPP). In contrast, the majority of mass-market autonomous and semi-autonomous deployments, including shared autonomous shuttles, last-mile delivery robots, and low-cost Level 2+ systems, continue to rely on single-antenna GNSS receivers aided by Satellite-Based Augmentation Systems (SBAS), which typically provide horizontal accuracy in the 1–4 m range and often degrade in dense urban environments [1,2]. When HD maps and high-end perception stacks are not available, these vehicles depend largely on GNSS-based localization, making them more vulnerable to noise, multipath, and map ambiguity—factors that directly affect their ability to identify the correct upcoming signal for SPaT delivery.

Among external information sources available to such vehicles, Signal Phase and Timing (SPaT) data play a critical role in ensuring safety at signalized intersections. Vision-based traffic light recognition remains vulnerable to occlusion, glare, and adverse weather [3,4,5], and many commercial systems treat SPaT as a redundant safety layer even when HD maps are available. For low-cost platforms that depend mainly on SBAS-aided GNSS and Standard Definition (SD) maps, it is difficult to guarantee reliable self-localization under all operating conditions, especially near complex intersections. As a result, an increasing number of architectures adopt a control-center or cloud-based component that performs map-matching on noisy GNSS trajectories and then selectively delivers SPaT messages to each vehicle. In such architectures, the quality of the infrastructure-side map-matching step directly determines the signal associated with each vehicle and thereby shapes the overall reliability of intersection-level safety services.

Despite this importance, infrastructure-side map-matching has received limited attention compared with onboard localization. Prior studies have examined geometric matching, topology-based search, or Kalman filter smoothing for low-cost GNSS, but most evaluations focus on onboard implementations or assume moderate network complexity [6,7,8,9,10,11]. Moreover, the literature rarely evaluates signal-matching accuracy explicitly, even though it is the primary performance indicator for vehicle-to-network (V2N)-based SPaT delivery in SD-map environments. There is also a lack of systematic evidence on how performance changes as road networks become denser and more interconnected, which is precisely where GNSS degradation and map ambiguity are most severe.

This gap motivates a systematic investigation of map-matching algorithms designed specifically for processing noisy GNSS trajectories at the control center. In particular, there is a need to understand how algorithmic design choices such as route constraints, Kalman filtering, geometric filtering, and spatial indexing affect link and signal association under varying GNSS errors and across road-network density levels. Accordingly, the aim of this study is to comparatively evaluate five representative infrastructure-side map-matching algorithms and to identify algorithmic structures that maintain robust link and signal matching for cost-constrained autonomous mobility services. To achieve this aim, the algorithms are tested under controlled GNSS perturbations using real-world trajectories from Jeju Island, South Korea, and a node–link-based SD road map that follows common digital map structures used in many countries. The comparative analysis provides new empirical evidence on the robustness of infrastructure-side map-matching under realistic GNSS conditions and identifies the design elements that most effectively support reliable SPaT delivery in low-cost autonomous vehicle deployments.

2. Related Work

2.1. Map-Matching Techniques in Autonomous Vehicles

Map-matching approaches for autonomous vehicles can be categorized into onboard and infrastructure-side methods. Onboard techniques dominate in premium fleets, where sub-decimeter accuracy is achieved through high-definition (HD) maps, multi-sensor fusion, and centimeter-level GNSS corrections such as RTK and PPP. Representative studies combine LiDAR point-cloud registration [12,13] and and deep-learning odometry [14] to construct highly accurate vehicle poses. However, these methods require costly sensors and continually updated HD maps, both of which remain unavailable for most mass-market autonomous deployments and are unlikely to reach broad adoption [15].

Low-cost platforms instead rely on SBAS-aided GNSS and Standard Definition (SD) maps, which typically provide 1–4 m accuracy [1,2]. Under these constraints, several lightweight onboard map-matching techniques have been explored, including geometric matching [7], topology-based search [8], and EKF-based smoothing [9]. Although effective in simple road structures, real-world evaluations consistently report notable degradation in link-matching performance in dense urban regions due to GNSS multipath and ambiguous road topologies [10,11]. These limitations motivate complementary infrastructure-side solutions capable of correcting noisy GNSS trajectories without requiring complex onboard sensing.

2.2. Signal Information Extraction and Its Relevance

Signal Phase and Timing (SPaT) information is one of the most safety-critical external data sources for intersection traversal [16,17]. Although vision-based traffic-light recognition continues to advance, real-world deployments still exhibit missed or unreliable detections under occlusion, glare, or adverse weather [3,4,5]. These persistent limitations have accelerated the adoption of cooperative mechanisms.

Roadside Unit (RSU) broadcasting via PC5 sidelink enables direct distribution of SPaT but is challenging to scale economically at a national level [18,19]. Selective V2N delivery over cellular networks has therefore gained traction [20], where the infrastructure transmits only the relevant signal information to each vehicle. Crucially, this process requires the infrastructure to determine the correct approach link and corresponding signal for every vehicle in real time. Thus, signal-matching accuracy becomes a core requirement of infrastructure-side map-matching, yet it remains largely underexamined in the existing literature.

Although 3GPP Releases 16 and 17 introduced PC5 unicast/groupcast and early NR Positioning [21,22], real-world field tests show that urban accuracy typically remains within the 5–10 m range [23,24]. Release 18 introduces carrier-phase enhancements and sidelink-assisted ranging, aiming to achieve centimeter-level NR positioning accuracy [24]. However, Release 18–class positioning capabilities are still in an early stage of commercial maturity, with hardware availability limited and real-world evaluations largely confined to controlled environments [25]. As a result, these features are unlikely to be widely deployed in cost-sensitive autonomous vehicle platforms in the near term. Accordingly, infrastructure-side map-matching remains a practical and necessary mechanism for ensuring reliable signal association in low-cost autonomous vehicles.

2.3. Limitations of Existing Infrastructure-Side Approaches

Infrastructure-side map-matching has historically received far less attention than onboard methods [6]. Traditional systems primarily employ geometric filtering [7], with EKF-based smoothing used as a deterministic enhancement [6,26]. Route constraints [27] and spatial indexing [28,29] have been studied individually, but the literature shows several notable gaps:

• The absence of systematic comparisons among geometric, route-based, Kalman-enhanced, and grid-based approaches under realistic low-cost GNSS conditions (1–5 m);
• Very limited attention to infrastructure-side map-matching, as most existing studies focus on onboard localization pipelines leveraging high-end sensors and HD maps;
• Limited studies that explicitly evaluate downstream signal-matching accuracy, despite its importance for selective V2N SPaT delivery;
• A lack of performance analyses across varying GNSS error levels and road-network densities, where GNSS accuracy degradation and map ambiguity become more pronounced;
• A scarcity of evaluations using continuous real-world driving trajectories, as many prior studies depend on small-scale test paths or synthetic traces rather than full-length road-network data.

These limitations are particularly consequential given current deployment trends. Commercial V2N systems either rely on onboard lane-level localization [30,31] or are beginning to adopt Release 18 positioning modules, which remain cost-prohibitive for most fleet-scale operations [21]. This situation underscores the need for robust, low-cost infrastructure-side map-matching solutions capable of supporting selective SPaT delivery under degraded GNSS conditions.

2.4. Contribution of This Study

This study addresses the above gaps by providing a comprehensive and systematic comparative evaluation of representative infrastructure-side map-matching strategies for low-cost autonomous vehicles operating with SBAS-aided GNSS and SD maps. While the individual algorithmic components, such as EKF-based refinement, route constraints, and geometric filtering, are well established in the literature, their relative effectiveness and robustness have not been previously analyzed within a unified infrastructure-side framework under realistic GNSS error levels and road-network densities.

Based on the comparative scope and problem setting considered in this study, it is observed that neither purely geometric matching nor map-matching approaches that combine geometric constraints with continuously applied filtering-based refinement are sufficient to ensure robust infrastructure-side operation across diverse GNSS conditions and road-network configurations. Geometric approaches are vulnerable to temporary GNSS degradation, while continuously filtered designs may propagate estimation errors when route ambiguity increases or candidate consistency is lost. These observations motivate a selective design in which route-level constraints are prioritized when reliable prior information is available, and filtering-based refinement is activated only as a fallback mechanism under degraded conditions.

In this context, this study examines a route–EKF-based map-matching strategy that integrates route-first filtering with EKF-based fallback as an operational design choice rather than a new theoretical construct. Unlike existing approaches that apply filtering in a fixed and continuous manner, the proposed strategy prioritizes route-level constraints when available and selectively activates EKF-based refinement only when route consistency cannot be maintained. This operational structure enables robust link and signal association under intermittent GNSS degradation while remaining suitable for real-time, infrastructure-side deployment under cost and scalability constraints. Through systematic experiments across multiple GNSS error levels and road-network densities, the presented evaluation identifies algorithmic structures that consistently maintain stable link and signal matching performance. Collectively, the results establish infrastructure-side map-matching as a practical and safety-relevant complement to PC5- and Release 18-based positioning capabilities in cost-constrained autonomous mobility deployments.

3. Methodology

3.1. System Architecture

Figure 1 illustrates the real-time map-matching system architecture developed in this study, which processes real-time vehicle state data and generates the corresponding signal-matching results required for safe autonomous driving. The architecture is divided into two major domains: the automated vehicle and the control center. On the vehicle side, the automated vehicle system contains on-board sensors such as LiDAR, radar, and cameras, along with the electronic control unit (ECU). These components provide raw sensor and state information through a dedicated vehicle-OBU interface. The on-board unit (OBU) collects essential autonomous driving data—including current latitude and longitude, heading, speed, longitudinal acceleration, and lateral acceleration—using its OBU Data Collector. The OBU Data Processor then combines these measurements with vehicle identifiers and timestamps and encodes them into a Protocol Buffer-based vehicle driving data message. The OBU Communication Module transmits this message to the control center.

Within the control center, the V2N server receives the uploaded message at the communication gateway and performs a Message Format Conversion step, transforming the Protocol Buffer payload into a standardized JSON message. The converted message is published to the vehicle driving data message topic of the Kafka-based message broker. The message broker provides real-time, reliable, and orderly delivery of vehicle data between autonomous vehicles and control centers, as commonly implemented in autonomous vehicle communication systems.

Once the driving data arrive at the broker topic, the Map-Matching Processor consumes the message through its Vehicle Driving Data Consumer. The processor executes two sequential engines: the Link-Matching Engine, which determines the road link on which the vehicle is currently located, and the Signal-Matching Engine, which identifies the traffic signals that the vehicle must receive at that position. The resulting link and signal information are forwarded to the Message Converter, where the Vehicle Driving Safety Message Composer reconstructs the output into the appropriate message format to be delivered back to the vehicle. This formatted message is published to the vehicle driving safety message topic. The V2N server subscribes to this topic, converts the message back into Protocol Buffer format, and transmits it to the OBU, where the OBU Message Receiver & Dispatcher delivers the final results to the vehicle.

In parallel, both the incoming vehicle driving data messages and the outgoing vehicle driving safety messages are ingested by the Data Saving Module. The module parses the messages into structured fields and stores them in the central database to support large-scale logging, monitoring, and offline analysis.

Although the overall architecture enables real-time closed-loop communication with operating vehicles, the Map-Matching Processor serves as the core analytic component responsible for determining both link and signal associations. Because different map-matching strategies exhibit varying levels of robustness depending on GPS noise and network density, it is necessary to compare multiple algorithmic approaches under controlled conditions. Therefore, this study evaluates five representative map-matching algorithms to determine which approach provides the most reliable performance under varying GPS error levels and road-network densities.

3.2. Integrated Framework for Map-Matching Evaluation

Based on the system architecture described in Section 3.1, an integrated evaluation framework was designed to assess the performance of the five map-matching algorithms under controlled and repeatable conditions. Figure 2 provides an overview of the overall evaluation workflow.

The framework follows the same end-to-end data flow used in the system architecture, from vehicle state generation through broker-based distribution to algorithmic processing. For evaluation purposes, however, the real vehicle, OBU, and V2N data paths were replaced with a driving-data emulator that can generate controlled input streams. The emulator produces time-stamped vehicle state records at one-second intervals and injects Gaussian noise with predefined standard deviations into the latitude and longitude fields, creating eleven GPS error groups for consistent and reproducible testing.

Each generated vehicle record was published to the vehicle driving data message topic, from which all five map-matching algorithms concurrently consumed identical input streams. Because the message broker delivers data in a consistent and orderly manner, each algorithm processed the same noisy trajectories without any interference or timing differences. The algorithms then produced structured outputs containing the matched link ID, matched signal ID, and associated driving metadata. For evaluation, these outputs were stored in a centralized database rather than being returned to the vehicle.

The stored results were subsequently analyzed using the evaluation metrics defined in Section 3.4.4 and the performance comparison procedures presented in Section 4. This framework provides a consistent and reproducible environment for evaluating and comparing the five map-matching algorithms.

3.3. Design and Mechanisms of Map-Matching Algorithms

This section is organized as follows. First, several algorithmic components used across the map-matching approaches, such as position refinement, adjacent-link candidate identification, and signal-selection mechanisms, are introduced. Subsequently, each algorithm is described in detail along with its corresponding pseudo-code to clearly present its operational logic.

3.3.1. Shared Components of the Algorithms

This subsection presents the components used across the map-matching algorithms. These include position refinement mechanisms applied in selected algorithms, geometric primitives for adjacent-link candidate identification, and signal-selection mechanisms used in multiple algorithms. Position refinement improves robustness under GPS noise, while the geometric primitives and signal-selection mechanisms support reliable link and signal association where applicable.

Extended Kalman Filter for Position Refinement (Algorithms 2, 4 and 5)

The vehicle state model considered in this study is inherently nonlinear, as it is defined in terms of latitude, longitude, and heading, while the available sensor inputs are limited to low-cost GNSS and basic inertial measurements. The EKF provides a practical and widely adopted solution for vehicle trajectory estimation in GNSS-based localization systems, and has been extensively used in autonomous driving and vehicle state estimation applications [32]. By supporting nonlinear state propagation with relatively low implementation complexity, the EKF offers stable real-time performance, making it well suited for scalable, infrastructure-side deployment. Algorithms 2, 4 and 5 refine raw GPS measurements using an EKF applied to the six-dimensional state vector

$${\mathbf{x}}_t={\left[\begin{array}{c}{\varphi }_t,{\lambda }_t,v_t,{\theta }_t,a_t^{\mathrm{lon}},a_t^{\mathrm{lat}}\end{array}\right]}^{\mathrm{T}},$$

where ${\varphi }_t$ and ${\lambda }_t$ denote latitude and longitude, $v_t$ the vehicle speed, ${\theta }_t$ the heading angle, and $a_t^{\mathrm{lon}}$ and $a_t^{\mathrm{lat}}$ the longitudinal and lateral accelerations.

For evaluation consistency, the EKF parameters were configured identically across Algorithms 2, 4 and 5. The measurement noise covariance R was set to reflect the overall magnitude of GNSS uncertainty represented by the predefined GPS error groups, while remaining fixed across all experiments. The process noise covariance Q was also fixed to model smooth vehicle motion and to avoid algorithm- or scenario-specific tuning effects.

No parameter optimization or scenario-dependent tuning was performed; a single EKF configuration was used throughout the study to ensure fair comparison and to isolate the impact of map-matching logic rather than filter parameterization. A detailed mathematical formulation of the EKF is provided in Appendix A for completeness.

Adjacent Link Candidates

Within a spatial buffer of $d_{\mathrm{buffer}}=15\mathrm{m}$ around the vehicle position $p_t=({\varphi }_t,{\lambda }_t)$, the set of nearby vertices is defined as

$${\mathcal{V}}_t=\{v\in V:d(p_t,v)\le d_{\mathrm{buffer}}\}.$$

Consecutive vertices form candidate links ℓ, each associated with a heading angle ${\theta }_{\mathcal{l}}$. A directional consistency filter is applied using

$${\mathcal{L}}_t^{\mathrm{head}}=\{\mathcal{l}:|{\theta }_t-{\theta }_{\mathcal{l}}|\le {45}^{\circ }\}.$$

Although the heading-based approach, later formalized as Algorithm 1, first sorts candidates by heading difference before evaluating distance, its final matched link still follows the same projection-based selection principle adopted in the other algorithms. The selected link is therefore expressed as first sorts candidates by heading difference before evaluating distance, its final matched link still follows the same projection-based selection principle adopted in the other algorithms. The selected link is therefore expressed as

$${\mathcal{l}}_t^{*}=arg\underset{\mathcal{l}\in {\mathcal{L}}_t^{\mathrm{head}}}{min}{d}_{\perp }(p_t,\mathcal{l}),$$

where ${\mathcal{l}}_t^{*}$ denotes the matched link at time t, and $argmin$ returns the link that minimizes the perpendicular projection distance.

Forward-Signal Selection (Algorithms 2–5)

For Algorithms 2–5, signal matching is performed only among signals mapped to the matched link ℓ. The cumulative distance from the upstream node to a signal s is denoted as $d_{\mathrm{sig}}^{\mathcal{l}}\left(s\right)$, and that to the vehicle position as $d_{\mathrm{veh}}^{\mathcal{l}}\left(t\right)$. A signal is considered forward-facing if it satisfies

$$d_{\mathrm{sig}}^{\mathcal{l}}\left(s\right)\ge d_{\mathrm{veh}}^{\mathcal{l}}\left(t\right).$$

Among all valid forward signals, the matched signal is chosen as

$$s_t^{*}=arg\underset{s\in {\mathcal{S}}_t^{\mathcal{l}}}{min}\left(d_{\mathrm{sig}}^{\mathcal{l}}\left(s\right)-d_{\mathrm{veh}}^{\mathcal{l}}\left(t\right)\right),$$

where $s_t^{*}$ denotes the selected signal associated with link ${\mathcal{l}}_t^{*}$.

3.3.2. Algorithm 1: Heading-Based Link and Signal Matching

Algorithm 1 performs link matching and signal association independently using only the vehicle’s raw GPS position and heading. Adjacent link candidates are retrieved using the geometric primitives described earlier. Among candidates that satisfy the directional condition $|{\theta }_t-{\theta }_{\mathcal{l}}| \le {45}^{\circ }$, the algorithm ranks links primarily by heading deviation and secondarily by distance, selecting the one most consistent with the vehicle’s current direction of travel.

Signal selection is conducted separately using a heading-based criterion. For each signal within a speed-dependent search radius,

$$d_{\mathrm{sig}}=max(10v_t,140\mathrm{m}),$$

the heading angle from the vehicle to the signal is computed. Valid signals satisfy

$$|{\theta }_t-{\theta }_s|\le {45}^{\circ }.$$

Among these, the nearest signal with the smallest heading deviation is selected as the matched signal.

Algorithm 1: Heading-Based Map Matching

Input: Position $p_t$, heading ${\theta }_t$, speed $v_t$, network G, signals S Output: Matched link ${\mathcal{l}}_t^{*}$ and signal $s_t^{*}$ Candidates ←ADJ_LINKS($p_t$, ${\theta }_t$, G); ${\mathcal{l}}_t^{*}\leftarrow$ RANK_BY_HEADING_AND_DIST (Candidates); SigRange $\leftarrow max(10v_t,140)$; SigCand ←SIGNALS_WITHIN($p_t$, SigRange); ValidSig ←FILTER_BY_HEADING (SigCand, ${\theta }_t$); $s_t^{*}\leftarrow$NEAREST_SIGNAL (ValidSig) return$({\mathcal{l}}_t^{*},s_t^{*})$

3.3.3. Algorithm 2: EKF-Based Map Matching

Algorithm 2 improves link-matching robustness by applying an EKF to refine the vehicle’s position and heading before performing matching. Using the six-dimensional state vector defined in Section 3.3, the EKF generates the corrected position ${\widehat{p}}_t$ and heading ${\widehat{\theta }}_t$. Link matching then proceeds using the common geometric primitives with ${\widehat{p}}_t$ and ${\widehat{\theta }}_t$.

Signal selection follows the offset-based forward-search rule described in the common primitives.

Algorithm 2: EKF-Based Map Matching

3.3.4. Algorithm 3: Route-Constrained Map Matching

Algorithm 3 restricts link- and signal-matching operations to a predefined route $R=\{L_1,L_2,\dots ,L_n\}$ associated with the vehicle. If no valid route is found, the algorithm immediately terminates. All geometric queries operate only within the links belonging to R. This constraint enhances matching stability on known trajectories and prevents off-route link switching.

Signal matching again follows the offset-based rule defined in the primitives.

Algorithm 3: Route-Constrained Map Matching

3.3.5. Algorithm 4: Grid-Based EKF Map Matching

Algorithm 4 reduces computational load by partitioning the network into $1\mathrm{km}\times 1\mathrm{km}$ grid cells. The EKF-corrected position ${\widehat{p}}_t$ is used to identify the active grid cell $G_{ij}$. Only links within this grid are considered for matching, enabling efficient geometric filtering without degrading accuracy.

Algorithm 4: Grid-Based EKF Map Matching

3.3.6. Algorithm 5: Hybrid Route–EKF Map Matching

Algorithm 5 integrates the strengths of Algorithms 2 and 3 by combining route-first matching with an EKF-based fallback mechanism. The algorithm first attempts to match links within the predefined route R. If no valid route-based match exists, it refines the position using an EKF update and performs a full-network matching. This hybrid design improves robustness in cases of temporary route loss or severe GPS drift.

Algorithm 5: Hybrid Route–EKF Map Matching

3.4. Experimental Setup

This section describes the experimental configuration established to evaluate the performance of the five map-matching algorithms under varying GPS errors and road network complexities. All algorithms were executed under identical data, hardware, and parameter conditions to ensure fairness and reproducibility.

3.4.1. GPS Error Simulation

The selected GPS error ranges (0–5 m) reflect realistic horizontal positioning errors reported for SBAS-aided, single-frequency GNSS receivers used in low-cost autonomous vehicles. SBAS-based systems typically provide meter-level positioning accuracy under nominal conditions, while field experiments indicate that GNSS positioning errors commonly remain within the 1–4 m range in both suburban and urban driving environments, with larger errors occasionally observed under severe multipath and signal obstruction conditions [1,2]. Accordingly, the maximum simulated error was capped at 5 m to focus the evaluation on realistic and practically relevant operating conditions, rather than rare worst-case GNSS failures.

Based on this error range, Gaussian noise was applied to the latitude and longitude coordinates of the ground-truth trajectories to simulate GPS signal fluctuations in a controlled and repeatable manner. Following empirical evidence from Specht (2021) [33], which demonstrated that latitude error standard deviations ($s_{\varphi }$) generally exceed longitude error standard deviations ($s_{\lambda }$) by approximately 25–39%, eleven distinct GPS error groups were defined, as summarized in

Table 1. Definition of GPS error standard deviation groups used for the simulation.

Group	1	2	3	4	5	6	7	8	9	10	11
GPS Error Standard Deviation (m) (Latitude, Longitude)	0.0, 0.0	0.5, 0.4	1.0, 0.8	1.5, 1.2	2.0, 1.6	2.5, 2.0	3.0, 2.4	3.5, 2.8	4.0, 3.2	4.5, 3.6	5.0, 4.0

.

Latitude standard deviations were increased from 0.0 m to 5.0 m in increments of 0.5 m, while longitude standard deviations ranged from 0.0 m to 4.0 m in increments of 0.4 m, maintaining the proportional difference consistent with empirical ratios. This configuration effectively captures key GPS drift characteristics reported in prior studies, including directional asymmetry and positional bias. Each group was subjected to ten repeated trials with randomized noise seeds to ensure robust evaluation of algorithmic stability under stochastic error conditions.

3.4.2. Network Density Classification and Spatial Scope

To analyze the sensitivity of each algorithm to road network complexity, the test areas were classified into three levels of network density: low, medium, and high. Low-density areas refer to suburban or rural regions characterized by sparse intersections and long straight segments, whereas medium-density areas represent moderately connected arterial roads, and high-density areas correspond to dense urban regions with short links and complex intersections. The road network density was quantified as the total sum of link lengths contained within a 140 m radius centered on each vehicle position. The 140 m radius was determined based on the minimum signal-search radius employed in Algorithm 1, derived from the operational characteristics of the autonomous driving dataset collected on Jeju Island, Republic of Korea.

Jeju Island was selected as the experimental site because it provides a clearly defined and well-managed road infrastructure, diverse road geometries and traffic environments, and established facilities for C-ITS and autonomous driving tests. Moreover, its insular geography minimizes external interference and ensures a controlled environment for real-world experiments.

Considering the island’s maximum speed limit (80 km/h) and the average deceleration distance required for safe stopping at signal changes, a 140 m range was determined to be the minimum effective distance for vehicle signal perception and response. For Algorithm 1, which uses a minimum signal-search radius as the baseline for signal acquisition and allows adaptive expansion with vehicle speed, increasing the minimum radius enlarges the baseline search area and enables earlier signal perception, while also increasing the number of candidate signals in dense road networks. As the candidate set grows, the likelihood of incorrect signal association may also increase. Conversely, reducing the minimum radius limits candidate complexity but may delay signal acquisition, potentially leaving insufficient time for the vehicle to respond smoothly to upcoming signal changes. Accordingly, the selected 140 m radius represents a practically balanced choice for Algorithm 1, ensuring timely signal perception while limiting the risk of incorrect signal association. This radius provides a consistent spatial basis for network-density classification and enables fair quantitative comparison of algorithm performance under varying GNSS error levels and road-network complexities.

3.4.3. Ground Truth and Validation Method

Ground-truth link and signal identifiers were established through a semi-automated verification process that combined the adjacent link query and matching module with manual validation in QGIS. When multiple signal candidates were present within a single link, manual inspection was conducted to eliminate ambiguity. While manual inspection was used in this study to establish reliable ground truth for evaluation, it was applied as an offline verification step to resolve ambiguous cases and ensure data correctness, and is not part of routine real-time control-center operation.

Each vehicle driving record was assigned a unique index, ensuring a one-to-one correspondence between ground-truth data and emulator-generated trajectories. For each experimental case, the outputs from Algorithms 1–5 were directly compared with the ground-truth dataset using this common index. Link matching accuracy and signal matching accuracy were then computed by verifying whether the matched link ID and signal ID from each algorithm corresponded exactly to those in the ground-truth records. This validation process ensured that the resulting performance metrics accurately represented each algorithm’s capability, free from any errors caused by index misalignment or data mismatching.

3.4.4. Evaluation Metrics

To quantitatively evaluate the performance of each map-matching algorithm, two primary metrics were defined: Link Matching Accuracy and Signal Matching Accuracy. Link Matching Accuracy is defined as the proportion of correctly identified links among all evaluated trajectory points, while Signal Matching Accuracy refers to the proportion of correctly matched traffic signals among trajectory points that contain valid ground-truth signal identifiers. These metrics are computed as follows:

$$\mathrm{Link}\mathrm{Matching}\mathrm{Accuracy}={\displaystyle \frac{N_{\mathrm{correct}\mathrm{links}}}{N_{\mathrm{total}\mathrm{links}}}}\times 100,$$

(1)

$$\mathrm{Signal}\mathrm{Matching}\mathrm{Accuracy}={\displaystyle \frac{N_{\mathrm{correct}\mathrm{signals}}}{N_{\mathrm{valid}\mathrm{signals}}}}\times 100,$$

(2)

where $N_{\mathrm{correct}\mathrm{links}}$ denotes the number of correctly identified road links, $N_{\mathrm{total}\mathrm{links}}$ represents the total number of evaluated trajectory points, $N_{\mathrm{correct}\mathrm{signals}}$ denotes the number of correctly matched traffic signals, and $N_{\mathrm{valid}\mathrm{signals}}$ refers to the number of trajectory points with valid ground-truth signal identifiers.

Statistical indicators—maximum, quartiles (Q1, Q3), median, minimum, and mean—were then computed across ten repeated trials for each GPS error group to assess algorithmic stability. The resulting accuracy distributions were visualized using box plots, as discussed in Section 4.

3.4.5. Execution Environment for Comparative Evaluation

This subsection describes the execution environment used for controlled comparative evaluation, rather than a requirement for simultaneous operational deployment of all algorithms. All map-matching algorithms were implemented in Java and executed concurrently within a single virtual machine (VM) hosted on the Microsoft Azure cloud platform (Microsoft Corporation, Redmond, WA, USA). The VM was configured with a Standard B16als v2 specification (16 vCPUs, 32 GiB RAM) running Ubuntu 22.04 LTS, ensuring sufficient computational capacity to handle real-time processing across all five algorithms simultaneously. The cloud-based environment was adopted over a local deployment to provide scalable resource allocation, process isolation, and consistent performance monitoring. This configuration allowed for parallel algorithm execution under identical hardware and software conditions while maintaining fair resource distribution and reproducibility. To support real-time data transmission and processing, a Kafka-based message broker was configured within the same virtual machine. The vehicle driving data emulator, developed in Kotlin, was responsible for generating synthetic driving data every second and injecting Gaussian GPS noise into the latitude and longitude coordinates. These data streams were then published to designated Kafka topics, which were subscribed to by the five Java-based map-matching algorithms for real-time processing. After completing the matching process, each algorithm produced structured output messages—containing matched link IDs, signal IDs, and original driving data—and republished them to corresponding Kafka result topics. All Kafka services were deployed locally within the VM to minimize network latency and to maintain synchronized timing across the data generation and algorithm execution pipelines. This integrated cloud-based setup ensured stable runtime behavior, minimized network latency, and provided a reproducible execution environment for evaluating algorithmic performance under consistent operational conditions.

4. Results and Discussions

4.1. Performance Under Varying GPS Error Levels

This section evaluates the performance of each map-matching algorithm under varying GPS error conditions. Gaussian noise was injected into the ground-truth trajectories according to the eleven standard deviation groups defined in Table 1. These groups replicate realistic localization uncertainty and directional bias of GPS sensors. Both link and signal matching accuracies were analyzed to assess the robustness and stability of the algorithms as GPS errors increased.

4.1.1. Link Matching Accuracy

Figure 3 and

Table 2. Average link matching accuracy by algorithm and GPS error standard deviation groups (on-route).

	GPS Error Standard Deviation Group (See Table 1)
Algorithm	Group 1	Group 2	Group 3	Group 4	Group 5	Group 6	Group 7	Group 8	Group 9	Group 10	Group 11
Algorithm 1	0.99027	0.99030	0.98956	0.98863	0.98731	0.98555	0.98077	0.97920	0.97631	0.97000	0.96223
Algorithm 2	0.99049	0.98957	0.98884	0.98770	0.98640	0.98398	0.98092	0.97826	0.97527	0.96953	0.96244
Algorithm 3	0.99308	0.99286	0.99171	0.99083	0.98956	0.98779	0.98348	0.98182	0.97859	0.97223	0.96461
Algorithm 4	0.99049	0.98959	0.98889	0.98771	0.98640	0.98384	0.98084	0.97785	0.97453	0.96852	0.96103
Algorithm 5	0.99308	0.99286	0.99192	0.99106	0.98986	0.98791	0.98383	0.98237	0.97879	0.97275	0.96546

present the link matching accuracies for the five algorithms across the defined GPS error groups. Across all conditions, the accuracy values of every algorithm remained above 0.96, indicating reliable spatial association even under increased positional uncertainty. As GPS noise intensified from Group 1 to Group 11, accuracies gradually declined and slight increases in variance were observed, yet the overall stability of all methods remained high throughout the tests.

Algorithm 1 applies a heading-prioritized link selection, where candidate links are first filtered by directional similarity and then by spatial proximity. Despite this procedural simplicity, its overall accuracy remained comparable to the other algorithms, ranging from 0.990 under low-error conditions to 0.962 in the highest noise group. This suggests that direction-based filtering can effectively support link identification when heading observations are stable. Algorithm 2, incorporating Kalman filtering for positional correction, demonstrated a stable trend but recorded slightly lower averages, ranging from 0.990 to 0.962, reflecting its limitation in handling extreme noise without additional structural constraints. Algorithm 3, leveraging predefined route information, sustained high consistency across all groups, with accuracies above 0.964 even under severe perturbations. This emphasizes the advantage of topological priors in mitigating GPS drift and maintaining map continuity. Algorithm 4 combines grid partitioning with Kalman filtering to improve spatial search efficiency; however, its accuracy dropped from 0.990 to 0.961 as GPS noise increased, marking the lowest value among the five algorithms at the highest error level. This indicates that while grid-based indexing accelerates candidate retrieval, it does not fully compensate for severe positional drift in high-noise environments. Notably, Algorithm 5, the hybrid route–Kalman algorithm, consistently achieved accuracies within the top-performing range across all groups, maintaining values above 0.993 in low-error scenarios and above 0.965 even at the largest deviation level. This resilience indicates that combining route constraints with Kalman-based position correction provides robust link-matching performance under degraded GPS conditions.

Furthermore, the distribution shown in Figure 3 suggests that Algorithm 5 tends to exhibit relatively small variation across repeated trials, supporting its stability and consistency under stochastic localization noise.

4.1.2. Signal Matching Accuracy

Figure 4 and

Table 3. Average signal matching accuracy by algorithm and GPS error standard deviation groups (on-route).

	GPS Error Standard Deviation Group (See Table 1)
Algorithm	Group 1	Group 2	Group 3	Group 4	Group 5	Group 6	Group 7	Group 8	Group 9	Group 10	Group 11
Algorithm 1	0.82593	0.82557	0.82473	0.82455	0.82390	0.82443	0.82302	0.82317	0.82222	0.82094	0.82056
Algorithm 2	0.99365	0.99283	0.99179	0.99112	0.98921	0.98793	0.98503	0.98262	0.97889	0.97383	0.96863
Algorithm 3	0.99394	0.99376	0.99284	0.99232	0.99060	0.98879	0.98551	0.98310	0.98093	0.97444	0.96894
Algorithm 4	0.99365	0.99285	0.99185	0.99114	0.98924	0.98782	0.98494	0.98230	0.97830	0.97294	0.96730
Algorithm 5	0.99394	0.99376	0.99302	0.99246	0.99090	0.98892	0.98578	0.98358	0.98113	0.97488	0.96950

present the signal matching accuracies for the five algorithms across the same GPS error groups. While Algorithms 2–5 exhibit a trend generally consistent with link matching, Algorithm 1 shows a distinct pattern because it relies solely on vehicle heading rather than distance-based association. Moreover, as defined in the evaluation metrics, the signal matching accuracy was calculated only at positions where ground-truth signals exist; therefore, it should be interpreted independently from link-matching results.

Across all test conditions, Algorithms 2–5 maintained accuracy values above 0.96, while Algorithm 1 stayed around 0.82 with little variation across groups. Unlike the other algorithms that leverage positional and filter-based corrections, Algorithm 1 depends mainly on heading consistency for signal association. Its accuracy remained nearly constant across all groups (0.82593–0.82056) but was the lowest overall. This confirms that a heading-driven geometric approach alone cannot ensure reliable signal association under increased GPS deviation. Algorithm 2, which applies Kalman filtering to correct GPS drift, achieved high accuracy under low noise levels, recording about 0.99365 in Group 1 and decreasing moderately to 0.96863 in Group 11. Algorithm 3, using predefined route information to limit the search space, sustained accuracies above 0.968 across all groups, demonstrating strong resilience against growing GPS noise. Algorithm 4 combines grid partitioning with Kalman filtering to improve spatial search efficiency; however, its accuracy decreased from 0.99365 to 0.96730 as GPS noise increased, showing that grid-based indexing can help organize candidate searches but offers limited benefit in mitigating severe positional drift. Algorithm 5, which integrates route constraints with Kalman-based position correction, consistently remained among the highest-performing methods across all groups, maintaining values above 0.993 under low-error conditions and above 0.969 even at the largest deviation. Overall, these results suggest that Algorithm 5, which first applies route constraints for signal association and then performs Kalman-based correction only when no valid match exists within the route, provides stable and robust signal-to-map association under degraded localization accuracy. The visual distributions in Figure 4 reveal that Algorithm 5 tended to exhibit relatively small variance across repeated trials but also formed part of a distinct high-performance cluster together with Algorithm 3. Algorithms 2 and 4, both employing Kalman-based correction, show nearly identical distributions and moderate accuracy declines as GPS noise increases, whereas Algorithms 3 and 5, which apply route-level constraints, maintain consistently higher accuracies and narrower variance ranges across all error groups. This finding suggests that route-constrained designs are generally effective in mitigating GPS noise accumulation compared with methods relying solely on Kalman-based correction, while the hybrid Algorithm 5 combines both mechanisms to achieve consistently strong and stable performance overall.

4.2. Performance Across Network Densities

This section evaluates the sensitivity of each map-matching algorithm to variations in road-network complexity. Link- and signal-matching accuracies were analyzed across three network density levels: low (0–1000 m), medium (1000–2000 m), and high (2000 m and above). Network density was defined as the total length of all road links within a 140 m radius around each vehicle position, consistent with the criterion described in Section 3.4.2. For every density level, eleven GPS error groups were applied to the test trajectories, and the mean accuracies of the five algorithms were obtained from ten repeated experiments to evaluate robustness and consistency.

The analysis considers link-matching performance across different road geometries, including all road types, curved segments, and straight segments, followed by an examination of signal-matching performance under the same network-density conditions. The detailed results and discussion are presented in Section 4.2.1 and Section 4.2.2.

4.2.1. Link Matching Accuracy Across Network Densities

The comparative evaluation of link-matching performance across the three network-density groups shows that accuracy declines as GPS noise increases, and this decline becomes more pronounced as road networks grow denser and more interconnected. Link-matching accuracy refers to the proportion of vehicle positions at which the predicted link corresponds exactly to the ground-truth link. High-density regions contain short links, clustered intersections, and frequent heading transitions; these characteristics cause even moderate localization errors to generate multiple plausible link candidates, increasing the likelihood of adjacent-link or lateral misassignments. In contrast, low-density environments with long and sparsely connected links inherently limit such ambiguities. These overall tendencies are clearly summarized in the heatmap of density-averaged accuracies (Figure 5). Algorithm 5 shows consistently strong mean performance across all three density levels, with mean accuracies of 0.9887 in low density, 0.9839 in medium density, and 0.9777 in high density. The remaining algorithms show slightly lower but still stable averages, ranging from 0.986 to 0.988 in low density and from 0.974 to 0.977 in high density. The heatmap confirms that, although all algorithms maintain high accuracy overall, increasing network density systematically magnifies the effects of GPS drift and widens performance differences among methods.

Across all road types, the GPS-error curves in Figure 6 reveal a consistent downward trend from Group 1 to Group 11. In low-density networks, accuracies begin around 0.994–0.996 in Group 1 and gradually decrease to approximately 0.971–0.973 by Group 11. This pattern reflects the stabilizing influence of long links with limited branching, which reduces the likelihood of confusing adjacent links. Medium-density networks show initial values near 0.987–0.990, but the performance gap among algorithms becomes clearer after Group 5, where latitude noise reaches 2.0 m and longitude noise reaches 1.6 m. Algorithm 4 shows the steepest reduction in this region, most likely because GPS drift can shift the estimated position across the 1 km × 1 km grid boundaries, so that a neighboring cell is selected and link candidates are drawn from an inappropriate spatial subset. High-density networks exhibit the sharpest declines, with accuracies falling from 0.988 to 0.992 in Group 1 to 0.945–0.947 in Group 11. Among all methods, Algorithm 5 records the smallest end-to-end loss, demonstrating that selectively combining route-based constraints with Kalman correction produces a more stable trajectory interpretation when the number of plausible link candidates increases. This finding suggests that hybrid priors—not distance or heading cues alone—are particularly effective for mitigating error propagation in compact urban road grids.

Curved road segments exhibit more moderate accuracy degradation compared with the overall trend, largely because continuous heading variation along a curve provides stronger directional cues that help distinguish adjacent link candidates even under GPS drift. This geometric characteristic enhances the separability of adjacent links, reducing the likelihood of misassignment at curve transitions. As shown in Figure 7, low-density curved networks begin with accuracies exceeding 0.997 in Group 1 and maintain values above 0.990 through Group 6. Even in Group 11, accuracies remain within 0.984–0.986, indicating that curvature mitigates the ambiguity typically induced by increased positional noise. Medium-density curved regions show a similarly stable initial trend, beginning around 0.987–0.991 and declining to approximately 0.963–0.968 in Group 11. This reduction is steeper than in low-density areas but still less severe than the drop observed in straight segments. In high-density curved networks, accuracies fall from approximately 0.986 in Group 1 to 0.953–0.960 at Group 11. The slightly irregular pattern near the final error groups is attributable to the relatively small number of high-density curved samples, which causes minor oscillations in the plotted trajectories. Among the algorithms, Algorithm 1 records the highest final accuracy within high-density curves, outperforming Algorithm 5 by roughly 0.003. This behavior highlights the advantage of heading-prioritized filtering in environments where the natural curvature of the road reduces front–back ambiguity more effectively than distance-only or Kalman-based updates. However, this advantage is limited to specific geometric conditions, and Algorithm 5 maintains strong overall performance across densities due to its hybrid combination of route constraints and selective Kalman correction. The curved-road findings confirm that geometry plays a critical role in moderating the impact of GPS drift, and that algorithms leveraging directional cues benefit more in non-linear path structures.

Straight road segments show the largest decline in link-matching accuracy among all road types. The primary reason is that consecutive straight links share almost identical heading values and similar geometric shapes, making it difficult to determine whether a vehicle is on the current link or the immediately following link when longitudinal GPS noise is introduced. Even small drift along the travel direction produces adjacent-link confusion, in which the predicted link shifts prematurely to the next segment or remains incorrectly on the previous one. This mechanism becomes more prominent as GPS errors increase, leading to a sharper degradation than that observed in curved environments, where natural geometric changes provide additional cues that help differentiate neighboring links. In low-density straight networks, this issue is further amplified by a localized U-turn maneuver included in the dataset. During the U-turn, vehicles remain on a long, uninterrupted straight link while executing a tight directional reversal. When GPS noise is added, many of these points are displaced toward parallel or opposite-direction links, resulting in a large cluster of unmatched or misclassified cases. This single pattern accounts for 6252 out of 8453 mismatches (approximately 74%) in the low-density straight category, driving much of the steep decline observed in this region. As shown in Figure 8, accuracies in low-density straight areas begin around 0.986–0.991 in Group 1 but fall sharply to 0.942–0.944 by Group 11. Medium-density straight roads exhibit a smoother decline, decreasing from roughly 0.984–0.988 to 0.954–0.956 across the same error range. High-density straight segments follow a comparable pattern, starting near 0.988–0.992 at Group 1 and dropping to approximately 0.944–0.947 in the highest error group. Unlike the low-density case, the medium- and high-density datasets do not include U-turn maneuvers. The performance degradation in these environments is therefore attributed primarily to longitudinal adjacent-link confusion rather than localized trajectory patterns. Furthermore, compared with curved segments—where geometric curvature helps separate sequential links—straight-road geometries provide no directional cues to counteract this GPS-induced ambiguity, leading to relatively larger drops. Across all straight-road conditions, Algorithms 3 and 5 demonstrate the most stable performance because their route-continuity constraints restrict abrupt transitions to neighboring links. Algorithm 5 consistently yields the highest accuracy under severe noise, indicating that its fallback Kalman update effectively suppresses the longitudinal link-switching errors that dominate straight-segment misclassifications, particularly in the U-turn region.

4.2.2. Signal Matching Accuracy Across Network Densities

The evaluation of signal-matching performance across the three network-density groups shows trends that are broadly consistent with link matching, while also reflecting the different definition of the signal-matching metric. Signal-matching accuracy is computed only at positions where at least one ground-truth signal exists and is defined as the proportion of those positions at which the predicted signal corresponds exactly to the ground-truth signal. As a result, the denominator for signal matching is smaller than that for link matching and is concentrated around conflict points such as intersections and stop lines. In dense urban networks, multiple signals may lie within a similar distance and orientation relative to the vehicle, so even small localization errors or offset miscalculations can cause the algorithm to associate the vehicle with an incorrect signal along the same or a neighboring link. In contrast, low-density environments tend to have fewer, more isolated signals, which naturally limit such ambiguities.

These overall characteristics are summarized in the heatmap of density-averaged signal-matching accuracies (Figure 9). Algorithms 2–5, which all rely on link-based association and offset comparison rather than pure heading, maintain high mean accuracies across all densities, typically above 0.985 in low-density networks and around the 0.978 range in high-density networks. Algorithm 5 again attains the highest mean values across the three density levels, while Algorithms 2 and 4 perform particularly well in denser networks where Kalman-based position correction reduces offset noise around intersections. Algorithm 1, which selects signals using radial distance and heading without explicit link constraints, shows markedly lower averages and exhibits limited sensitivity to network density, confirming that heading-only association is insufficient under complex signal layouts. The heatmap confirms that these four link-based algorithms (Algorithms 2–5) maintain similarly high accuracy across densities, and that signal matching is less sensitive to algorithmic design than link matching because the association is evaluated only at positions where ground-truth signals exist.

Across all road types, the GPS-error curves in Figure 10 show a monotonic decline from Group 1 to Group 11, with clear separation between Algorithm 1 and the link-based Algorithms 2–5. In low-density networks, Algorithms 2–5 start in the high 0.992–0.984 range in Group 1 and decrease gradually to around 0.970–0.971 by Group 11, indicating that sparse signal layouts and long approach links reduce the impact of positional noise on signal association. In medium-density networks, initial accuracies remain close to those of low density but begin to diverge more noticeably as GPS noise increases. Algorithm 2, which incorporate Kalman filtering before computing link offsets, preserve slightly higher accuracies in the mid to high error groups than Algorithms 3 and 5, suggesting that temporal smoothing is beneficial when multiple signals are distributed along intersecting approaches. High-density networks exhibit the largest overall reductions, with Algorithms 2–5 decreasing to the 0.944–0.946 range by Group 11. Even in this challenging setting, Algorithm 5 maintains the most balanced performance across error levels, while Algorithm 1 remains around 0.95 regardless of density or GPS group, underscoring the limitations of purely heading-driven signal selection without link context.

Curved road segments, shown in Figure 11, exhibit a more gradual decline in signal-matching accuracy compared with straight segments. This behavior is largely attributable to the link-matching characteristics of curved geometries. As discussed in Section 4.2.1, continuous heading changes along a curve provide strong directional cues that help distinguish adjacent links even under GPS drift. Because signal association is performed only after a link has been correctly identified, the improved separability of neighboring links directly contributes to higher signal-matching stability on curved roads. In low-density curved regions, signal accuracies remain high across all algorithms, beginning around 0.994–0.995 in Group 1 and remaining above 0.985 by Group 11. Medium-density curves show slightly larger reductions, decreasing from approximately 0.993–0.994 at Group 1 to about 0.968–0.974 in the highest error group. High-density curved networks exhibit the sharpest drops within this road type, with values falling from roughly 0.979 at Group 1 to 0.952–0.960 at Group 11. Across all curved-road conditions, Algorithms 2 and 4 show slightly higher accuracies in medium- and high-density areas due to their Kalman-based smoothing, which helps stabilize heading variations and maintain more reliable link associations under increased GPS noise. Algorithm 5 also maintains strong and consistent performance across all curved densities, although the Kalman-related advantage of Algorithms 2 and 4 becomes marginally more visible when the road geometry contains prolonged curvature and the density of surrounding links increases. Overall, these results indicate that the geometric continuity of curved road segments facilitates more stable link identification, which in turn supports more reliable signal association compared with straight-road environments where adjacent-link ambiguity is more severe.

Straight road segments, shown in Figure 12, exhibit more pronounced signal-accuracy degradation than curved segments as GPS noise increases. On straight approaches, multiple signals often lie along the same alignment, and GPS drift can shift the estimated position forward or backward along the corridor, causing the associated stop-line to be confused with the next or previous signal. This effect mirrors the adjacent-link ambiguity observed in straight-segment link matching and becomes more influential as road networks grow denser. In low-density straight networks, Algorithms 2–5 begin in the 0.988–0.991 range at Group 1 but decline to approximately 0.941–0.942 by Group 11. These baseline values remain lower than those of curved segments because straight approaches provide fewer geometric cues for distinguishing among longitudinally aligned signal candidates. Medium-density straight regions start near 0.987–0.988 and decrease steadily to approximately 0.955–0.956 as GPS noise intensifies. High-density straight segments follow a similar trend but terminate slightly lower, converging to the 0.943–0.946 range at the highest error group. Across straight-road conditions, Algorithms 3 and 5 benefit from route-continuity constraints, which narrow the candidate signals to those associated with links on the predefined path and thereby reduce confusion among closely spaced stop lines. However, Algorithms 2 and 4 remain competitive when GPS errors are large, since Kalman-based smoothing limits abrupt jumps in the estimated stop-line position and stabilizes offset computations along the corridor. Overall, Algorithm 5 delivers the most robust performance across densities and geometries by combining route-based filtering with selective Kalman correction, while Algorithms 2 and 4 provide particularly strong signal-matching robustness in dense, curved environments where temporal smoothing of the trajectory is most effective.

4.3. Implications for Real-World Deployment

The findings of this study have several implications for the design and deployment of V2X-based signal information services, particularly in cost-constrained autonomous fleets. First, the consistently high link- and signal-matching accuracies observed for Algorithms 2–5, even under severe GPS perturbations and high network densities, indicate that infrastructure-side map-matching can reliably support SPaT dissemination with only SBAS-aided GNSS and SD maps. This is especially relevant for shared autonomous shuttles, last-mile delivery robots, and low-cost Level 2+ systems that cannot accommodate HD maps or high-end perception stacks but still require dependable intersection-level safety support.

Second, the comparative results provide practical guidance on algorithm selection for different operational contexts. The hybrid route–EKF algorithm (Algorithm 5) offers the most robust performance across all GPS error levels and density regimes, making it a strong candidate for fleets operating predominantly on predefined service corridors or dispatch routes where route information is readily available. In contrast, Algorithm 2 and Algorithm 4, which emphasize Kalman-based smoothing and grid-based indexing, respectively, offer attractive trade-offs between robustness and computational efficiency for services with more flexible routing or for large-scale deployments where scalability of candidate search is critical. These observations suggest that infrastructure operators can tailor the map-matching strategy to their service patterns, balancing robustness, prior information, and processing cost. It should be emphasized that the comparative conclusions in this study are drawn from robustness trends observed under increasing GNSS uncertainty and network density, rather than from marginal accuracy differences at a single noise level. In particular, we avoid claiming a statistically decisive “winner” based on sub-percent accuracy gaps in the lowest-noise regime. Moreover, because map-matching results are generated from continuous vehicle trajectories, consecutive samples are not independent. Applying point-wise hypothesis tests that treat each timestamp as an independent trial may therefore exaggerate statistical significance. For deployment-oriented decision making, we therefore interpret the repeated-run distributions (ten runs per condition) as an empirical characterization of stability, focusing on consistency under stress conditions where candidate ambiguity grows and operational failures are more likely.

Third, the experimental architecture demonstrates that all five algorithms can be executed concurrently in a cloud-hosted environment using a Kafka-based message broker while maintaining real-time operation at one-second update intervals. This confirms that the proposed framework can be integrated into existing control centers without specialized hardware and can scale to multiple vehicles and intersections by adjusting cloud resources. By decoupling map-matching from the vehicle and executing it at the control center, the framework also simplifies software updates and algorithm upgrades, enabling operators to progressively adopt more advanced methods such as Algorithm 5 without modifying onboard software. Overall, the results support the view that infrastructure-side map-matching is a practical and scalable complement to PC5-based broadcasting and emerging NR positioning solutions in connected-intersection deployments.

4.4. Limitations and Future Work

The limitations of this study arise from three perspectives: the characteristics of the data used for evaluation, the scope of the algorithms examined, and the assumptions made at the system level. From a data perspective, although the experiments were performed using real autonomous-vehicle trajectories, GNSS errors were generated by injecting Gaussian noise rather than collected under live SBAS or dense-urban GNSS conditions. This approach provides controlled and repeatable perturbations but does not capture non-Gaussian effects such as persistent multipath bias, deep-canyon signal blockage, or short-duration outages. In addition, all trajectories were collected on Jeju Island, which limits the diversity of road geometries included in the analysis. The study also relied on an SD node–link map without lane-level or semantic features, and therefore assumes that the map is locally accurate and complete.

From an algorithmic perspective, the five methods evaluated in this work rely primarily on GPS position, vehicle heading, and EKF-refined inertial measurements. Advanced sensing modalities such as LiDAR, camera-based landmarks, wheel odometry, or HD-map semantics were not incorporated, and the evaluation does not include probabilistic or learning-based map-matching techniques that may better adapt to non-Gaussian noise characteristics or dynamically changing network densities. While recent data-driven and learning-based approaches have demonstrated promising performance under noisy GNSS conditions, their reliance on large-scale labeled training data, limited transferability across regions, and increased computational and operational costs pose practical challenges for infrastructure-side deployment in cost-constrained V2N settings.

At the system level, the Kafka-based V2N communication pipeline was evaluated under idealized conditions with timely and lossless message delivery. Real-world uncertainties such as transmission delays, temporary network congestion, or occasional message loss were not considered, and these factors may influence the responsiveness and reliability of infrastructure-side map-matching when deployed at scale. From an evaluation perspective, this study focuses on comparative robustness trends observed across repeated runs under increasing GNSS uncertainty and network density. Building on this framework, future work can further formalize trajectory-aware statistical evaluation by aggregating link- and signal-matching outcomes at higher-level units, such as per-trajectory or intersection-approach segments, to enable more rigorous significance analysis while respecting temporal dependence in vehicle trajectories.

These limitations suggest several directions for future research. One direction is to evaluate the algorithms in additional cities and more diverse road environments to examine their robustness under different geometric layouts and GNSS conditions. Another important direction for future research is to extend the current framework to include probabilistic or learning-based map-matching methods that can adapt to non-Gaussian GPS errors and heterogeneous road-network structures. The Kafka-based architecture used in this study also enables future experiments that explicitly analyze end-to-end latency, throughput, and message-loss scenarios, providing deeper insight into real-world operational reliability. Finally, future work may explore hybrid designs that combine infrastructure-side map-matching with lightweight onboard priors, as well as the incorporation of emerging NR-positioning signals and semantic map cues, to support more accurate lane-level SPaT association while maintaining the cost advantages of SD-map-based autonomous mobility systems.

5. Conclusions

This study conducted a comprehensive comparative evaluation of five infrastructure-side map-matching algorithms for low-cost autonomous vehicles operating under diverse GNSS error levels and road-network densities. By combining real-world trajectories from Jeju Island with controlled GPS perturbations, the analysis demonstrated how algorithmic structures such as route constraints, Kalman filtering, geometric filtering, and spatial indexing influence link- and signal-matching accuracy under realistic localization uncertainty. The results confirm that infrastructure-side map-matching can reliably support cloud-assisted signal information services in V2N-based architectures built on SD maps and GNSS measurements.

Across all experiments, the hybrid route–EKF algorithm (Algorithm 5) consistently exhibited strong and stable performance across a wide range of GNSS error levels and network-density conditions. Its link accuracy ranged from 0.99308 to 0.96546 across GPS error groups and from 0.9887 to 0.9777 across network-density groups, while signal accuracy ranged from 0.99394 to 0.96950 and from 0.9865 to 0.9790 in the same conditions. These results suggest that route-level constraints, when supplemented by selective EKF-based refinement, can enhance resilience against lateral and longitudinal drift in dense or geometrically ambiguous environments. Other route- and Kalman-based approaches also demonstrated robust performance, whereas the heading-based method exhibited limitations in reliably distinguishing forward-facing signals. The real-time operation achieved on the Kafka-based V2N pipeline demonstrates the feasibility of infrastructure-side map-matching in a cloud environment. Future studies should further evaluate scalability under multi-vehicle and multi-intersection conditions and examine latency, throughput, and message reliability in realistic V2N communication scenarios to fully assess operational robustness. Although the evaluation relied on controlled noise models and trajectories from a single geographic region, the findings provide a solid foundation for extending the framework to more diverse urban environments and for integrating probabilistic or learning-based map-matching strategies.

Collectively, this study offers empirical evidence and practical insights that support the development of reliable and scalable infrastructure-side traffic-signal information services—including MAP and SPaT message delivery—for cost-constrained autonomous vehicle deployments.