Coverage Without New Sensors: How Far Can Virtual Traffic Sensing Move the Data?

TL;DR for operators

A traffic model can give an agency observations on roads where no detector exists, but pushing those observations farther across the network does not automatically make the resulting dataset more useful.

Guastella et al. propose a simulation-driven augmentation method that assigns each physical traffic sensor to a distinct unsensored road edge with similar simulated traffic behavior.¹ In Brussels, the shortest tested candidate range, one to five directed road-edge steps, produced the lowest errors. Under the paper’s strongest reported parameter setting for that range, RMSE reached 50.1 for one calibration strategy and 58.2 for the other. Repeated synthetic experiments in Namur also favored the one-to-five-hop range.

For operators, the relevant use case is bounded: extending analytical coverage onto uninstrumented corridors when installing hardware is slow or uneconomic. The virtual readings are not new measurements. They are simulated traffic states already implied by the underlying model, relocated to other road edges. Calibration errors therefore travel with them.

The paper provides good methodological evidence that where simulated observations are relocated matters. It does not yet show that the augmented data improve forecasting, routing, control, or other downstream decisions, and the spatial displacement mechanism is not a demonstrated privacy guarantee.

More distance can make the augmentation worse

Suppose an agency has dependable counts on several major corridors but needs data for nearby roads without detectors. A calibrated traffic simulation already estimates traffic throughout the network. The tempting move is to extract simulated observations from unsensored locations and treat the broader geographic spread as an augmentation of the original dataset.

The Brussels results show why displacement itself is the wrong objective.

Across the tested configurations, candidate locations one to five directed road-edge steps from the physical sensor produced the lowest average RMSE: 70.34 under the paper’s “simfree” calibration and 96.81 under its “simbased” calibration. Expanding the allowed range to one to ten hops raised those values to 96.87 and 131.19. Excluding the nearest candidates entirely, with a three-to-ten-hop range, raised them again to 101.46 and 138.83.

The implication is narrower than “keep virtual sensors close.” Distance is functioning as a constraint on behavioral resemblance. Once nearby edges that share similar flows are removed from consideration, geographic novelty can be purchased at the cost of traffic fidelity.

One reporting detail deserves care. In the simfree baseline comparison, the proposed method scores 101.46 RMSE for the three-to-ten-hop range, versus 87.67 for QR-pivot placement and 83.48 for random placement, so the proposed rule does lose its advantage there. The corresponding simbased row, however, reports 138.83 for the proposed method, lower than both QR-pivot at 173.81 and random at 147.45. The source package’s narrative characterization of both baselines outperforming the method under this range is therefore not fully consistent with its own table.

That inconsistency does not change the stronger sensitivity result: moving candidate locations farther away consistently increases the proposed method’s own reconstruction error.

The method searches for a surrogate, not a second measurement

The paper calls the simulated observation on an unsensored edge a virtual sensor. It is not hardware. It is a road location selected to stand in for a physical sensor while providing a displaced observation derived from simulation.

For each physically sensored road edge, the method first searches downstream through the road graph. Candidate edges must satisfy two conditions: simulated traffic must plausibly continue toward them, and their directed distance must fall inside a configured hop interval.

Candidates are then scored using two signals:

The first term rewards simulated vehicle-flow continuity from the original edge toward the candidate. The second penalizes differences in a simulated traffic metric such as occupancy, speed, or travel time. The weight $\alpha$ controls the balance.

After scoring, a global greedy assignment creates an injective matching: different physical sensors receive different virtual edges rather than several sensors collapsing onto the same surrogate location.

The augmented count itself is straightforward:

In other words, the replacement observation is simply the simulated vehicle count recorded at the assigned virtual edge during the same hour.

That definition sets the evidence boundary. The algorithm does not infer another independent real-world count from historical data. It exposes another part of the network state already produced by the simulation.

Modest traffic-state regularization works better than either extreme

The Brussels sensitivity tests help explain what makes a good surrogate.

For the one-to-five-hop range, the strongest tested setting uses occupancy with $\alpha=0.2$. It produces MAE/RMSE of 14.6/50.1 under simfree calibration and 14.5/58.2 under simbased calibration. The selected virtual edges are only about 1.86 and 1.92 hops away on average.

Neither extreme performs as well. When $\alpha=0$, the scoring effectively removes the turn-continuity reward; when $\alpha=1$, it removes the traffic-metric penalty. The best result instead comes from retaining both signals while giving substantial weight to traffic-state similarity.

This is best interpreted as an ablation and sensitivity result rather than a universal tuning rule. Occupancy at $\alpha=0.2$ is strongest in the tested Brussels configuration, but Namur shows some variation by metric and demand. What transfers more consistently is the preference for the shorter candidate range.

The repeated Namur experiments reinforce that point. Across synthetic traffic demand of 50,000, 100,000, and 150,000 vehicles, with 30 runs per configuration, the one-to-five-hop interval produces the lowest reported RMAE for every listed metric-demand comparison. For example, using speed, the best RMAE is 0.231 at 50,000 vehicles and 0.228 at 100,000, both with $\alpha=0.2$. Error grows sharply when the search range expands or nearby candidates are excluded.

Those runs are primarily a robustness test: they show that the local-surrogate preference is not confined to one Brussels calibration. They do not establish performance under arbitrary cities, network conversions, or real demand shifts.

Aggregate similarity is necessary but not sufficient

The augmented Brussels data reproduce the broad morning and evening demand peaks and closely preserve the network-wide average speed profile. That is useful evidence that the procedure does not destroy the dominant temporal structure of the dataset.

It is also an incomplete validation target.

The paper’s sensor- and region-level analyses show why. Errors at individual locations can be substantially larger even when positive and negative deviations cancel in aggregate. A citywide hourly curve can therefore look convincing while particular corridors remain poor surrogates.

For an agency using the method, validation should consequently follow the decision unit. If the intended use is network-level exploratory planning, aggregate agreement may carry meaningful weight. If the output will inform corridor-specific interventions, local count errors become the relevant acceptance criterion.

The paper does not test that second step against a downstream operational task.

The business case is temporary analytical coverage, not replacement evidence

What the paper directly shows: a calibrated or synthetic traffic model can be used to relocate sensor observations onto unsensored edges while retaining substantial spatiotemporal resemblance, provided candidate locations remain sufficiently local and behaviorally similar.

Cognaptus inference: a transportation agency could use this mechanism when it needs analytical coverage before additional physical infrastructure is available. Examples include exploratory work on uninstrumented corridors, maintaining a consistent development dataset after sensor-network changes, or testing planning scenarios while empirical collection catches up.

The decision boundary should follow the quality of the simulation. A virtual reading produced from a strongly calibrated corridor and used for exploratory analysis deserves more authority than one generated from a poorly constrained part of the network and fed directly into an automated operational decision.

This also changes the ROI question. The comparison is not simply “virtual sensor versus physical sensor.” A more realistic decision is whether the cost of improving simulation calibration plus local validation is lower than the cost and delay of installing hardware for the particular analytical need.

The paper does not measure that economic trade-off.

Calibration quality is the hard boundary

The method inherits whatever the traffic model gets wrong.

Traffic calibration is underdetermined: different vehicle trajectories can reproduce the same observed counts. The simulations also omit realistic traffic-light programs because public signal data were unavailable, and OpenStreetMap-to-SUMO conversion errors are not corrected. Virtual sensing can redistribute the model’s inferred information spatially, but it cannot resolve ambiguities that the model never resolved in the first place.

Three additional boundaries remain open.

First, Brussels uses real counts from 369 sensors but only for April 3, 2024. Second, no experiment tests whether augmented observations improve forecasting, learning, routing, or control. Third, although relocating observations may appear to create spatial separation from physical sensors, the paper provides no formal analysis showing that the hop-distance mechanism reduces trajectory-reconstruction or other privacy risks.

Those are not peripheral qualifications. They determine how much operational authority the augmented data should receive.

Extend coverage only as far as similarity survives

The paper’s strongest contribution is not that simulation can manufacture more traffic data. It cannot.

It shows a disciplined way to reuse a calibrated simulation’s network-wide state at additional road locations, with explicit controls over traffic continuity, behavioral similarity, displacement, and one-to-one assignment.

The experiments then identify the constraint that matters most: coverage remains useful only while the surrogate still resembles the traffic behavior it is meant to represent. Across both cities, nearby candidates perform better than wider displacement. In Brussels, pushing the search away from nearby edges can even erase the method’s advantage over simpler placement rules.

For transportation operators, virtual sensing is therefore best treated as a model-dependent extension of analytical reach. It may defer some immediate sensing needs and make sparse datasets more spatially usable. It should not be mistaken for additional empirical evidence, nor granted more decision authority than the simulation underneath it can justify.

Cognaptus: Automate the Present, Incubate the Future.