Source-linked AI summary
Simulation-Driven Vehicular Traffic Data Augmentation: Extending Sensor Coverage Through Virtual Sensing
Davide Andrea Guastella, Eladio Montero Porras, Evangelos Pournaras, Gianluca Bontempi
TL;DR
Sparse urban sensor networks limit the spatial diversity of traffic data and make datasets obsolete when infrastructure changes. This paper uses simulation-based virtual sensors selected by a graph-search heuristic, reducing RMSE by approximately 59% relative to random and QR-pivot placement while preserving spatio-temporal traffic patterns.
Problem
Sparse urban sensor coverage limits spatially diverse traffic datasets, while infrastructure changes can make collected training data partially obsolete and costly to replace.
Method
The framework generates virtual sensor measurements from calibrated or synthetic traffic models using graph search, similarity scoring, and minimum spatial displacement.
Results
Approximately 59% lower RMSE than random and QR-pivot placement was achieved for the recommended hop range, while preserving spatio-temporal traffic patterns.
Takeaways & Limitations
The method extends traffic information to unmonitored road segments while preserving observed sensor-level and regional traffic dynamics.
Takeaways & Limitations
Augmented-data quality is bounded by the underlying traffic model, so calibration errors may propagate into virtual sensor observations.
Abstract
from arXiv · showhide
Urban traffic management relies on sensor networks whose spatial coverage is limited by deployment costs and privacy regulations. Machine learning models trained on such sparse data cannot generalize to unmonitored locations and must be retrained whenever the sensor infrastructure changes. We propose a simulation-based methodology that addresses this problem by generating augmented traffic count datasets in which each physical sensor is replaced by a virtual sensor placed at a surrogate location in the road network. Virtual sensors are selected by a graph-search heuristic that jointly maximises vehicle-flow continuity and traffic-metric similarity between the original and surrogate locations, while enforcing a minimum spatial displacement to ensure diversity of observed traffic conditions. We validate the method on two Belgian cities: Brussels, using a calibrated model, and Namur, using synthetic models. The augmented datasets preserve the bimodal daily demand profile and the dynamics of traffic at the observed locations.
1 Introduction
Sparse, costly sensor networks leave many roads unmonitored and make fixed-location training data difficult to generalize or replace. The paper proposes simulation-based virtual sensing to expand spatial coverage while preserving traffic behavior and sensor correspondence.
- Motivation: Urban sensor networks cover only part of the road network because installation and maintenance costs limit deployment.Many secondary and residential streets remain uninstrumented despite their relevance to traffic modeling and management.
- Problem: Models trained on fixed sensor locations learn location-specific patterns, while infrastructure changes can make collected training data partially obsolete.Recollecting representative data after relocation, road construction, or privacy-driven camera removal can be prohibitively expensive.
- Proposed framework: The framework generates augmented traffic datasets from calibrated or synthetic traffic models while preserving the original sensor network’s structure.It uses turn counts and edge-level traffic statistics to generate realistic measurements at virtual locations.
- Methodological scope: Random virtual placements may produce unrepresentative traffic data, while purely data-driven generators lack observations for unmonitored edges.The paper establishes a general methodology rather than evaluating benefits for a specific prediction or control application.
- Virtual sensing: For each physical sensor, the method selects a nearby surrogate with similar volume, speed, occupancy, and travel-time behavior, creating a one-to-one virtual correspondence.Surrogates are selected within a limited number of topological hops to improve spatial coverage while maintaining statistical consistency.
2 Background
The background frames synthetic data augmentation as a response to limited or sensitive training data and reviews traffic-sensor placement and reconstruction methods. It distinguishes this work’s network-level surrogate-sensor formulation from prior local reconstruction and deployment-optimization approaches.
- Data augmentation: Data augmentation and synthetic-data generation address insufficient training data and can expand small, biased, or sensitive datasets.Synthetic data augmentation has been applied in healthcare and finance, including diagnostic classification and fraud detection.
- Traffic sensor placement: Recent traffic research combines sensor placement with machine-learning-based traffic reconstruction to address limited observability and sparse measurements.The reviewed Traffic Sensor Location Problem concerns selecting sensor numbers and locations under budget and coverage constraints.
- Traffic sensor placement: Traffic-sensor placement research increasingly emphasizes network observability, information gain, and scalable optimization.Existing approaches use algebraic and graph-theoretic methods or information-gradient frameworks to guide placement.
- Traffic data augmentation: Variational Autoencoders augment work-zone traffic-estimation data by learning a low-dimensional latent space for realistic synthetic samples.The method was evaluated on 212,000 hourly traffic-volume records from Utah work zones spanning 2016–2019.
- Traffic sensor placement: Pham et al. place traffic sensors by ranking links for reconstruction using matrix factorization with column pivoting and deep reinforcement learning.Their two-step method selects measurements for maximal reconstruction and applies a policy-gradient, one-dimensional convolutional model.
- Research gap: This work formulates sensor observation generation as a network-level surrogate task that preserves global traffic dynamics rather than interpolating locally.The passage identifies this formulation as not previously studied to the authors’ knowledge.
3 Proposed Method
The proposed method uses traffic simulation to extend sparse physical sensor coverage with virtual readings on uncovered road-network edges. It selects distinct surrogate edges through flow-aware graph search and a score combining vehicle-flow continuity with traffic-metric similarity.
- Method overview: Simulation generates virtual sensor readings on uncovered road-network edges, using either calibrated or synthetic traffic models as input.The resulting measurements extend the spatial coverage of datasets supplied to calibration methods.
- Method overview: The method constructs an injective mapping from each physical sensor edge to a distinct virtual sensor edge outside the physically sensored set.The TSLP heuristic processes every sensored edge and hourly interval before constructing the mapping.
- Candidate search: GetCandidates uses breadth-first search over directed, simulated traffic flows to identify candidate edges within the hop interval [hmin, hmax].Only edges with positive simulated turn counts are traversed, while the lower bound excludes locally redundant locations and the upper bound limits exploration.
- Surrogate scoring: Candidate quality combines turn-count similarity, which rewards vehicle-flow continuity, with traffic-metric similarity, which favors comparable simulated traffic conditions.The two terms are normalized by α ∈[0, 1], and metric differences are penalized.
- Surrogate assignment: A greedy matching strategy ranks valid physical-sensor–candidate pairs by descending time-averaged score and assigns candidates uniquely.Candidates must lie outside the physically sensored edge set.
4 Experimental Results
The experimental results evaluate the proposed traffic-count augmentation methodology in Brussels and Namur. The analysis compares ground-truth and augmented counts, examines spatiotemporal errors, and assesses preservation of network-wide traffic dynamics.
- Experimental setup: The experiments evaluate the proposed traffic-count augmentation methodology using Brussels and Namur as Belgian use cases.The section introduces the simulation tool, evaluation metrics, and mobility scenarios before presenting the method’s results.
- Dataset comparison: The evaluation compares traffic counts between the ground-truth and augmented datasets.
- Traffic dynamics: The results examine spatiotemporal error structure and preservation of network-wide traffic dynamics.
4.1 Simulation Tool
The study uses the open-source SUMO traffic simulator with a simplified microscopic configuration. SUMO supports microscopic and mesoscopic simulation, while the chosen configuration partially models vehicle behavior at intersections under right-of-way rules.
- The simulation uses the open-source traffic simulator SUMO.
- SUMO supports microscopic simulation of individual vehicles and their dynamics, as well as mesoscopic simulation using queues and coarse intersection models.
- The study configures SUMO with a simplified microscopic model that partially simulates vehicle behavior at intersections.Vehicles remain subject to right-of-way rules in this configuration.
4.2 Evaluation Metrics
The evaluation compares augmented traffic counts with ground-truth counts using MAE and RMSE. Together, these metrics characterize average deviations and sensitivity to large errors.
- Mean Absolute Error (MAE): MAE measures the average absolute difference between augmented and real vehicle counts for each sensor–hour pair.It is computed over n sensor–hour observations, with ˆy and y denoting augmented and real traffic counts.
- Root Mean Square Error (RMSE): RMSE is the square root of the mean squared difference between augmented and real traffic counts.By squaring residuals before averaging, RMSE penalizes large deviations more heavily than MAE.
- Metric interpretation: MAE and RMSE capture complementary aspects of augmentation quality.MAE reflects average absolute deviation, whereas RMSE emphasizes disproportionately large individual errors.
- Metric interpretation: Low MAE with high RMSE indicates accurate augmentation on average but occasional large outliers from mismatched virtual sensors.The interpretation concerns comparisons between augmented traffic counts Da and ground-truth counts Dr.
4.3 Mobility Scenarios
The evaluation covers Brussels and Namur using OSM-derived vehicular road networks, real 24-hour sensor data for Brussels, and synthetic traffic models for Namur. Namur’s scenarios reproduce bimodal urban demand and use repeated independent runs across vehicle volumes.
- Network preparation: Both city scenarios use OpenStreetMap topology converted for simulation and filtered to include only vehicular roads.The considered networks are Brussels and Namur.
- Brussels scenario: Brussels uses hourly vehicle counts from 369 real sensors collected over one 24-hour period on April 3, 2024.The data were provided by Brussels Mobility.
- Namur scenario: Namur uses synthetic models because empirical traffic data were unavailable, generating fixed vehicle volumes over 24 hours through stochastic regional-level path construction.Vehicle departures are allocated using the procedure described for the synthetic scenario.
- Namur scenario: The synthetic Namur demand follows a bimodal hourly profile with peaks assumed at 8AM and 5PM to represent morning and evening rush hours.The profile is designed to reflect realistic urban mobility patterns.
- Namur scenario: For each vehicle volume, the Namur scenario uses 30 independent runs so results are not tied to one random demand realization.Each resulting traffic model is evaluated independently with the proposed augmentation method.
4.4 Experimental Results (Brussels Scenario)
In the Brussels scenario, the proposed augmentation performs best with nearby candidate sensors, α=0.2, and the occupancy metric. The augmented data preserve daily demand and traffic dynamics, while localized boundary mismatches account for concentrated errors.
- Hop-range sensitivity: The simfree strategy consistently achieves lower RMSE than simbased, with both reaching minimum error under hop range [1, 5].Expanding to or increases RMSE monotonically for both strategies; simfree also shows higher standard deviation.
- Sensitivity to α and simulation metric: At α=0.2 with occupancy, MAE is 14.6 for simfree and 14.5 for simbased, while RMSE is 50.1 for simfree versus 58.2 for simbased.This configuration selects virtual sensors 1.86–1.92 hops away on average, the shortest across configurations.
- Hourly vehicle count profiles: Both augmented profiles reproduce Brussels’ bimodal demand, including morning 07:00–09:00 and evening 17:00–19:00 rush-hour peaks.The simbased profile tracks aggregate ground truth more closely, whereas simfree slightly underestimates peak-hour volume.
- Spatial and per-sensor errors: Approximately 65% of sensor–interval pairs have errors below 20 vehicles per hour, while a long tail reaches 700 vehicles per hour.The tail is associated with boundary sensors whose virtual counterparts face structurally different demand patterns and small candidate pools.
- Traffic dynamics: Input and augmented profiles remain closely aligned over 24 hours, including speed drops during morning and evening congestion peaks.This indicates placement on edges with similar vehicle volumes and consistent underlying traffic dynamics.
- Baseline comparison: For [1, 5], the proposed method reduces simfree RMSE by 59.2% relative to QR-pivot and by 58.9% relative to random placement.The corresponding RMSE pairs are 70.34 vs. 172.43 for QR-pivot and 70.34 vs. 170.99 for random placement.
4.5 Experimental Results (Namur Scenario)
Namur experiments use fully synthetic traffic demand to evaluate augmentation quality across congestion levels and repeated simulations. The proposed heuristic performs best with hop range [1, 5], maintaining stable accuracy and rapidly converging results across demand scenarios.
- Experimental setup: Namur evaluation uses entirely synthetic demand, enabling controlled traffic conditions and statistical assessment across repeated experiments.This contrasts with Brussels, which uses a calibrated model based on real sensor data as ground truth.
- Experimental setup: 30 independent simulations evaluate each combination of demand level, α, and simulation metric across 50,000, 100,000, and 150,000 vehicles.All simulations use the same virtual sensors, covering approximately 30% of the road network.
- Evaluation metric: RMAE is used as a scale-independent accuracy metric because the experiments compare different traffic demand levels.RMAE is defined over sensor–hour observations using augmented and real traffic counts.
- Augmentation results: The hop range [1, 5] consistently produces the lowest RMAE across all demand levels, whereas makes RMAE more than triple.Excluding nearest candidates removes the most similar edges and forces selection from edges with weaker traffic continuity; also has consistently low standard deviations.
- Robustness and convergence: RMAE remains stable from moderate to high congestion, and cumulative mean RMAE stabilizes within the first 15 experiments across all three demand levels.The cumulative mean is computed over runs using the best configuration identified for each demand level.
5 Discussion
The method reliably produces augmented traffic datasets that preserve urban mobility structure, provided near-neighbor search, limited metric regularization, and appropriate calibration are used. Its validity remains bounded by traffic-model quality, simulation simplifications, and unresolved privacy-analysis needs.
- 5 Discussion: Recommended settings are hmin=1, hmax≤5, and α=0.2, combining near-neighbor candidates with limited metric regularization for turn-count matching.These conditions preserve spatio-temporal urban-mobility structure at sensor and regional levels without distorting turn-count matching.
- 5 Discussion: Augmented-data quality is bounded by the underlying traffic model because calibration errors in simulated flows may propagate to virtual-sensor observations.The method extends existing traffic information rather than correcting inaccuracies in the input model.
- 5 Discussion: Missing realistic traffic-light programs and uncorrected OSM-to-SUMO conversion errors reduce simulation realism but do not undermine validation on disjoint sensor sets.The validation goal is to show that input and augmented datasets produce similar traffic patterns when evaluated at separate sensors.
- 5 Discussion: The configurable minimum hop distance controls displacement between physical and virtual sensors, but formal privacy analysis of trajectory reconstruction difficulty remains necessary.The current work does not quantify how difficult it is to reconstruct individual vehicle trajectories from augmented data.
- 5 Discussion: Although calibration uses sparse real-sensor counts, augmentation extracts additional virtual observations from the calibrated traffic model.Calibration estimates vehicle trajectories whose simulated counts resemble real sensor counts, while augmentation generates observations at virtual locations.
6 Conclusion and Future Work
The paper establishes a lightweight, deterministic simulation-based heuristic for augmenting traffic data through virtual sensor placement while preserving observed traffic dynamics. Experiments in calibrated Brussels and synthetic Namur show replicated spatio-temporal traffic patterns, while future work targets downstream task effects and improved placement strategies.
- Conclusion: The heuristic addresses data scarcity by evaluating virtual sensor locations through simulation-derived turn counts and traffic metrics, then applying greedy injective matching.It is computationally lightweight and requires no training data.
- Conclusion: Given ε=0, the method produces deterministic, reproducible output from any calibrated or synthetic simulation model.This supports reproducibility across the calibrated Brussels network and synthetic Namur scenarios.
- Conclusion: The augmented datasets replicate the original traffic counts’ spatio-temporal patterns at both sensor and spatial-region levels in Brussels and Namur.The evaluation covers a calibrated Brussels network and synthetic Namur scenarios.
- Conclusion: Approximately 59%: the proposed scoring function reduces RMSE relative to the baseline comparison.The passage reports the reduction as approximate and relative to the baseline comparison.
- Future Work: Future work will assess generated datasets for learning, forecasting, and control, while exploring alternative scoring and search strategies and joint real-plus-augmented data.The proposed placement improvements target sparse regions and wider hop constraints; combining real and augmented data may improve downstream predictive traffic models.