Fine-Grained Road Traffic Inference

• Fine-grained Road Traffic Inference (FRTI) is a framework that reconstructs detailed traffic states from sparse, coarse, or heterogeneous observations.
• It integrates per-vehicle sensing, road segment forecasting, and lane-level graph generation to enhance spatial and temporal resolution.
• Advanced methodologies, including Bayesian filters, tree ensembles, and graph neural networks, enforce physical constraints and improve model robustness.

Fine-grained Road Traffic Inference (FRTI) denotes a family of inference problems in which high-resolution traffic states are reconstructed, predicted, or characterized from observations that are sparse, indirect, coarse-grained, or heterogeneous. In recent literature, the term covers per-vehicle sensing and characterization from roadside telecommunication fibers, road-segment inundation risk estimation from traffic and crowdsourced signals, 5-minute segment-level speed prediction, coarse-to-fine temporal forecasting from 10–30 minute observations, road-network-guided magnification of coarse urban flow maps, 1-minute movement-level forecasting under signal control, and road-to-lane generation of lane-level traffic states (Liu et al., 2023, Yuan et al., 2021, Le et al., 2017, Li et al., 8 Jun 2026, Liu et al., 2021, Liu et al., 2023, Li et al., 25 Jul 2025). Across these formulations, the common objective is to increase spatial, temporal, or semantic granularity without proportionally increasing sensing density.

1. Scope, granularity, and task formulations

The literature does not use FRTI as a single standardized task. Instead, it appears as a set of related formulations distinguished by the target spatial unit, the temporal resolution, and the semantics of the inferred state. At the finest physical granularity, TelecomTM performs per-vehicle detection, direction-aware tracking, position and speed estimation, and characterization through wheelbase and weight inference from virtual strain sensors derived from telecommunication fiber-optic cables (Liu et al., 2023). At the finest network granularity, RoadDiff formalizes FRTI as a road-to-lane task with road-level spatio-temporal inputs $X_R \in \mathbb{R}^{|V_R| \times T \times d}$ and lane-level targets $Y_L \in \mathbb{R}^{|V_L| \times T \times k}$ (Li et al., 25 Jul 2025).

Other formulations operate at the road-link or movement level. The flooding study defines a binary inundation status for a road segment at a given time and outputs a probability of inundation at the road-link level, using fine-grained traffic data and crowdsourced reports as label proxies (Yuan et al., 2021). FDTI defines fine-grained traffic prediction at 1-minute resolution on a movement-level graph in which each node is a traffic movement and each edge corresponds to an admissible vehicle transition (Liu et al., 2023). Local Gaussian Processes and the distributed LSTM systems target detector- or segment-level short-term speed prediction at 5-minute granularity (Le et al., 2017, Lee et al., 2021).

A second axis of variation is whether the problem is predictive or reconstructive. STRP formulates cross-granularity forecasting from coarse histories to fine future sequences through Window-Based Fine-Grained Prediction (WBFP) and Duration-Based Fine-Grained Prediction (DBFP) (Li et al., 8 Jun 2026). RATFM and UrbanMSR treat the problem as fine-grained urban flow inference from coarse spatial observations under sum-preserving constraints, although their native formulation is grid-based rather than road-segment-based (Liu et al., 2021, Yuan et al., 2024). This suggests that “fine-grained” in FRTI refers not to a single output type, but to a systematic increase in resolution relative to the available measurements.

2. Observation regimes and sensing substrates

FRTI systems span a wide range of sensing substrates. TelecomTM converts redundant dark fibers into roadside virtual strain sensors using distributed acoustic sensing with $\phi$-OTDR. A coherent interrogator measures Rayleigh backscatter phase shifts quasi-linearly proportional to axial strain, producing a dense 1D strain field with 1 m channel spacing, 10 m gauge length, and up to 250 Hz sampling, with a single interrogator covering up to $\sim 100$ km (Liu et al., 2023). The sensed response contains a quasi-static component dominated by ground deformation due to vehicle weight and a dynamic component arising from wheel-road interactions at surface features.

Traffic-operations-oriented FRTI typically relies on fixed detectors or traffic feeds. Local GP models use static sensors and TMC-derived speed matrices at 5-minute bins (Le et al., 2017). DistPre and DistTune use per-detector speed series aggregated at 5-minute intervals, with each detector treated as an individual prediction locus and with normalization by a fixed factor such as a 70 mph speed limit for similarity assessment (Lee et al., 2020, Lee et al., 2021). FDTI augments movement-level traffic volumes with roadnet-enriched features $s_t^i = \{p_t^i, l^i, d^i\}$, where $p_t^i$ is green signal time, $l^i$ is movement length, and $d^i$ is movement direction (Liu et al., 2023).

Coarse-to-fine inference systems additionally incorporate structural side information. RATFM uses OpenStreetMap road polylines rasterized into a weighted road map $M_r = \overline{M}_r \otimes M_w$, where $M_w$ is derived from historical flow averages (Liu et al., 2021). UrbanMSR uses neighborhood-level and city-level geographic information for multi-scale representation learning (Yuan et al., 2024). FRTP extends the data model further by federating route search logs, toll data, and exogenous features such as weather, calendar, and events, while reconciling different sampling periods through a unified temporal alignment module (Ge et al., 2024).

A separate branch of FRTI uses traffic as an indicator of non-traffic phenomena. The flooding-risk study uses INRIX average speeds at 5-minute intervals and Waze hazard and closure reports to derive segment-level flood labels, then combines these with topographic, hydrologic, and temporal precipitation features (Yuan et al., 2021). In this setting, the traffic observation is not the final target but a proxy for road accessibility under extreme weather.

3. Inference architectures and learning paradigms

Methodologically, FRTI includes probabilistic state-space models, tree-based ensembles, Gaussian processes, recurrent models, graph neural architectures, transformer-based coarse-to-fine magnifiers, and diffusion models. TelecomTM uses a spatial-domain Bayesian filter and Rauch–Tung–Striebel smoothing over the state $Y_L \in \mathbb{R}^{|V_L| \times T \times k}$0, where $Y_L \in \mathbb{R}^{|V_L| \times T \times k}$1 is arrival time at a virtual sensor and $Y_L \in \mathbb{R}^{|V_L| \times T \times k}$2 is time per unit distance. Its motion model

$Y_L \in \mathbb{R}^{|V_L| \times T \times k}$3

directly exploits sensor adjacency along the road and Newtonian motion constraints (Liu et al., 2023). The resulting smoothed arrival times yield position and speed, with $Y_L \in \mathbb{R}^{|V_L| \times T \times k}$4 or $Y_L \in \mathbb{R}^{|V_L| \times T \times k}$5.

For road inundation inference, the dominant models are tree ensembles. The flooding study trains Random Forest and AdaBoost models in scikit-learn, with Random Forest configured at 200 trees and max depth 90. It produces segment-level probabilities $Y_L \in \mathbb{R}^{|V_L| \times T \times k}$6 and reports repeated 5-fold cross-validation over balanced datasets (Yuan et al., 2021). In this formulation, FRTI is a segment-time classification problem with engineered features rather than a direct spatio-temporal dynamical model.

Speed-prediction formulations emphasize scalability. Local GP models factorize the speed matrix $Y_L \in \mathbb{R}^{|V_L| \times T \times k}$7 via non-negative matrix factorization to localize correlated subsets in space and time, then train a Gaussian process only on the corresponding local set (Le et al., 2017). DistPre and DistTune instead customize per-detector LSTMs by tuning $Y_L \in \mathbb{R}^{|V_L| \times T \times k}$8 with the Nelder–Mead method, while allowing model sharing across detectors whose normalized patterns satisfy an AARD threshold (Lee et al., 2020, Lee et al., 2021). These systems make scalability a first-class modeling constraint rather than treating it as a post hoc engineering concern.

Coarse-to-fine forecasting introduces architectures that explicitly manage granularity mismatch. STRP couples Tree Convolution for hierarchical spatial aggregation with Inverse Dilated Convolution for temporal extrapolation from coarse inputs to fine outputs, with the granularity ratio $Y_L \in \mathbb{R}^{|V_L| \times T \times k}$9 controlling the dilation stride (Li et al., 8 Jun 2026). RATFM uses a multi-directional 1D convolution to encode road geometry and a transformer whose queries are road features, thereby separating short-range road-relative distribution modeling from long-range road-aware dependency capture (Liu et al., 2021). FDTI constructs a Fine-grained Traffic Spatio-Temporal Graph with signal-gated dynamic edges and applies Dynamic Mobility Convolution, followed by explicit inflow and outflow prediction under traffic flow conservation (Liu et al., 2023).

Lane-level FRTI has recently been framed as a spatio-temporal graph node generation problem. RoadDiff uses a Road–Lane Correlation Autoencoder–Decoder to generate initial lane states from road embeddings, then refines them with a Lane Diffusion Module whose forward process conditions lane corruption on road features and whose reverse process is regularized by road-to-lane physical constraints (Li et al., 25 Jul 2025). This marks a shift from deterministic magnification toward generative lane-state inference.

4. Structural priors, physical constraints, and interpretability

A defining feature of FRTI is that increased granularity is rarely inferred without structural priors. TelecomTM requires per-sensor geo-localization, per-sensor pattern classification, and transmissibility calibration to correct for polarity flips, heterogeneous gains, conduit changes, and fiber slack or spooling. Its Bayesian fusion stage then uses multi-sensor spatial continuity and Newtonian motion constraints to suppress large and complex noise (Liu et al., 2023). The underlying claim is not merely statistical correlation but physically constrained trajectory recovery.

In road-network and flow-magnification settings, conservation and topology play the same role. UrbanMSR formalizes coarse-to-fine inference with an aggregation matrix $\phi$0 such that $\phi$1, and its $\phi$2-normalization guarantees that fine-grained predictions sum to the observed coarse flows (Yuan et al., 2024). STRP similarly formulates coarse signals as aggregations of fine signals over a ratio $\phi$3, and it explicitly notes an aggregation consistency term as a practical extension, although the reported experiments use only the fine-grained MSE objective (Li et al., 8 Jun 2026). RATFM does not impose hard conservation for inflow and outflow datasets, but it does encode explicit road priors through a weighted road map and road-guided transformer queries (Liu et al., 2021).

Signal timing becomes a first-order prior at 1-minute resolution. FDTI argues that fine-grained traffic is non-smooth between nodes because green and red phases rapidly change the effective downstream mobility. Its dynamic edge weight

$\phi$4

for valid transitions formalizes signal-controlled mobility, and its next-step volume update

$\phi$5

enforces a discrete conservation law (Liu et al., 2023).

Lane-level inference adds explicit road-to-lane constraints. RoadDiff enforces that road speed equals the average of its lane speeds and road flow equals the sum of its lane flows:

$\phi$6

These constraints are used during training and reverse denoising to maintain physically plausible lane states (Li et al., 25 Jul 2025). A recurrent misconception is that fine-grained inference is equivalent to unconstrained super-resolution; the literature instead shows that physical, topological, and aggregation constraints are central to stable performance.

5. Empirical settings and reported performance

Reported results are strongly task-dependent. The same label “fine-grained” covers detection accuracy, RMSE, MAE, MAPE, percent error, ROC-AUC, and end-to-end customization time, depending on whether the system targets vehicles, road segments, detectors, movements, or lanes. The table summarizes representative results exactly as reported in the cited studies.

System	Granularity and setting	Reported results
TelecomTM (Liu et al., 2023)	Per-vehicle inference on a two-way, two-lane urban road with 1120 virtual sensors	90.18% vehicle detection accuracy; position MAE reduced from 140.09 m to 5.16 m; speed MAE reduced from 17.18 km/h to 3.57 km/h; wheelbase $\phi$7 percent error; weight $\phi$8 percent error
Flooding-risk RF (Yuan et al., 2021)	Segment-level inundation risk with INRIX and Waze labels	Harvey: RF AUC 0.860, accuracy 0.900; Imelda: RF AUC 0.790, accuracy 0.764
DistTune (Lee et al., 2021)	5-minute per-detector speed prediction on 110 I5-N detectors	Total time 3144 min $\phi$9 56 min from Case 1 to Case 4; 31 models for 110 detectors; average AARE with full customization $\sim 100$0; sharing cases still $\sim 100$1
STRP_att (Li et al., 8 Jun 2026)	Coarse-to-fine forecasting with $\sim 100$2 min	WBFP MAE: METR-LA 5.22, PeMS-Bay 2.80; DBFP MAE: METR-LA 7.02, PeMS-Bay 3.08
RATFM (Liu et al., 2021)	Road-network-guided coarse-to-fine urban flow inference	XiAn: RMSE 16.260, MAE 5.746, MAPE 21.48%; ChengDu: RMSE 21.248, MAE 8.033, MAPE 18.65%
RoadDiff (Li et al., 25 Jul 2025)	Road-to-lane inference of speed and flow	PeMS speed: MAE $\sim 100$3 vs best deep baseline $\sim 100$4; HuaNan speed: MAE $\sim 100$5 vs $\sim 100$6
FDTI (Liu et al., 2023)	1-minute movement-level forecasting	Nanchang H=1: RMSE 1.30, MAPE 6.55; Manhattan H=1: RMSE 1.20, MAPE 4.84; Hangzhou H=1: RMSE 1.46, MAPE 7.20

These results are not directly comparable across rows because the targets and metrics differ. Even within a single formulation, performance depends sharply on the sensing regime and topology. TelecomTM reports that the outbound lane nearer to the fiber has 4.7× lower position MAE and 1.5× lower speed MAE than the farther inbound lane, while FDTI shows that 1-minute movement-level data are intrinsically non-smooth and that over-smoothing is a major failure mode of conventional graph predictors (Liu et al., 2023, Liu et al., 2023). Likewise, the flooding study finds that performance is more stable for Hurricane Harvey than for Tropical Storm Imelda because the label source and event hydrometeorology differ (Yuan et al., 2021).

6. Deployment regimes, robustness, and open issues

FRTI deployment choices reflect the dominant bottleneck of each formulation. TelecomTM emphasizes ubiquitous coverage, low maintenance, privacy preservation, and the absence of pavement cuts or cameras, with per-vehicle real-time throughput at 250 Hz and $\sim 100$7 sensors described as feasible on commodity hardware (Liu et al., 2023). DistPre and DistTune treat large or growing detector networks as the core systems problem, using master–worker clusters and model sharing to avoid redundant per-detector training (Lee et al., 2020, Lee et al., 2021). FRTP addresses a different operational constraint: raw route search logs, toll data, and exogenous variables may be distributed across data owners, so it uses federated optimization with shared backbones and local adapters to preserve data locality (Ge et al., 2024).

Robustness issues are equally heterogeneous. TelecomTM must address environmental changes, nearby construction, fiber layout heterogeneity, polarity flips, and crosstalk, and it reports that single-sensor true positive rate drops by 77% during crosstalk events while multi-sensor fusion lifts it from 12% to 54% (Liu et al., 2023). Flood-risk inference faces label noise and domain shift because INRIX null speeds and Waze crowd reports encode different operational notions of flooding (Yuan et al., 2021). STRP notes that large granularity ratios and extreme nonstationarity make fine reconstruction difficult, and that without a consistency term the aggregate of fine predictions may drift from coarse constraints if both are available (Li et al., 8 Jun 2026). RoadDiff identifies merges, diverges, and mis-specified road-to-lane correspondences as sources of systematic bias (Li et al., 25 Jul 2025).

Several active directions recur across the literature. These include uncertainty quantification for lane-level or cross-granularity outputs, stronger event-aware covariates, dynamic or multi-fiber integration, adaptive diffusion schedules, factor-graph or particle smoothing for complex maneuvers, and more explicit treatment of data governance through federated or privacy-preserving training (Liu et al., 2023, Ge et al., 2024, Li et al., 25 Jul 2025). A plausible implication is that future FRTI systems will combine three ingredients rather than optimize only one: physically grounded constraints, heterogeneous sensing fusion, and deployment-aware learning architectures.