🤖 machine learning

Simulation-Free Estimation of Traffic Flows from Sparse Count Data

This paper proposes a simulation-free, weighted least-squares optimization method that estimates time-varying traffic flows from sparse aggregated sensor data by partitioning a study area into regions and deriving edge-level trajectories, demonstrating superior accuracy and computational efficiency over baseline methods on the Brussels road network.

Original authors: Davide Guastella, Gianluca Bontempi

Published 2026-06-23

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Davide Guastella, Gianluca Bontempi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to figure out exactly how many cars are driving on every single street in a busy city like Brussels. The problem is, you only have a few "watchtowers" (traffic sensors) placed on a tiny fraction of the roads. You can see the cars passing these watchtowers, but you have no idea what's happening on the thousands of streets in between.

Usually, to solve this puzzle, city planners would run a giant, complex computer simulation. They would guess a traffic pattern, run the simulation, see if it matches the watchtowers, tweak the guess, and run it again. They'd do this hundreds of times until it finally worked. But this is like trying to find a needle in a haystack by building a new haystack every time you check—it takes forever and requires massive computing power.

The New Approach: A Two-Step "Smart Guess"

The authors of this paper propose a much faster, "simulation-free" way to solve this puzzle. Instead of running a full simulation, they use a clever two-step math trick to reconstruct the traffic flow.

Step 1: The "Neighborhood" View (The Big Picture)
First, they ignore the individual streets and divide the city into large "neighborhoods" (regions). They ask a simple question: "How many cars moved from Neighborhood A to Neighborhood B?"

They use a mathematical optimization tool (think of it as a super-smart calculator) to figure out the best number of cars to assign to the paths connecting these neighborhoods. This calculator has a special rule: it only trusts the paths that pass near the actual watchtowers. If a path goes through a dark alley with no sensors, the calculator gives it a lower score. This ensures the solution stays grounded in the real data they actually have.

Step 2: The "Street-Level" View (The Details)
Once they know how many cars are moving between neighborhoods, they need to figure out exactly which streets those cars took.

Imagine you know 100 cars left the "Downtown" neighborhood and arrived in the "Suburbs." Now, you have to guess which specific streets they used. The method looks at all possible street routes between those two areas. It scores each route based on two things:

Volume: Does this route pass by enough sensors to explain the total number of cars we saw?
Timing: Do the cars on this route show up at the same times as the cars we saw at the sensors? (e.g., if the sensors saw a rush hour spike at 8:00 AM, the route must also show a spike at 8:00 AM).

The method picks the best-matching streets and assigns the cars to them, creating a detailed map of individual car journeys without ever running a traffic simulation.

Why is this a big deal?

Speed: The old simulation method took about 69 minutes to figure out the traffic pattern for a whole day. This new method did it in about 23 seconds. That's roughly 180 times faster.
Accuracy: Even though it's faster, it was just as good at matching the real traffic counts as the slow simulation method. In fact, it beat a standard tool called "RouteSampler" by a huge margin.
No "Crutches": Many other methods need a "cheat sheet" (a pre-existing map of where people usually go, called an OD matrix) or GPS data from thousands of cars. This method works with only the sparse sensor data, making it useful for cities that don't have perfect data.

The Catch (Limitations)

The authors are honest about the limitations. Because they are working with incomplete data (sparse sensors), the puzzle is technically "under-determined." This means there isn't just one single correct answer; there are many possible traffic patterns that could fit the sensor data.

The "Blind Spot" Problem: If a part of the city has no sensors, the method has to guess. It might guess the cars took the shortest path, which might not be what actually happened if there was traffic.
Over-fitting: If the computer runs the math for too long, it starts to "overthink." It might put all the cars on the specific streets that have sensors just to make the numbers match perfectly, ignoring the fact that real traffic spreads out more.

In Summary

This paper presents a new, lightning-fast way to reconstruct city traffic maps using only a few scattered sensors. Instead of running slow, heavy simulations, it uses a smart, two-step math process: first figuring out the big flow between neighborhoods, and then filling in the specific street details based on sensor timing and volume. It's like solving a jigsaw puzzle by looking at the edge pieces and the picture on the box, rather than trying to force every piece into place by trial and error.

Technical Summary: Simulation-Free Estimation of Traffic Flows from Sparse Count Data

Problem Statement

The central challenge addressed is the estimation of time-varying traffic flow patterns and edge-level trajectories from sparse, aggregated vehicle counts collected by fixed sensors (e.g., induction loops, cameras). This problem is intrinsically under-determined, as multiple trajectory configurations can reproduce the same aggregate counts.

Existing approaches face significant trade-offs:

Simulation-based methods achieve high accuracy by iteratively adjusting demand parameters until simulated counts match observations. However, they require running full traffic simulations at every iteration, leading to prohibitive computational costs that hinder scalability for city-level networks.
Offline/Analytical methods are faster but often rely on strong assumptions (e.g., free-flow conditions, predefined route choice models), require prior Origin-Destination (OD) matrices, or depend on dense sensor coverage and trajectory data (e.g., GPS probes) to constrain the solution space.

The paper asks whether accurate, time-varying traffic flows can be estimated from sparse sensor counts alone, without simulation feedback or prior OD matrices, at a computational cost scalable to city-level networks.

Methodology

The proposed method is a simulation-free, hierarchical framework that reconstructs traffic flows in two distinct stages: Region-Level Assignment and Edge-Level Assignment.

1. Region-Level Assignment (Global Optimization)

The study area is partitioned into spatial regions ( $R$ ). Instead of optimizing at the edge level immediately, the method solves a global optimization problem to assign vehicle flows to feasible region-to-region routes ( $\Phi$ ).

Contribution Matrix ( $C$ ): A weighted contribution matrix encodes sensor coverage. Each entry $C(\phi, \rho)$ represents the weighted proportion of sensors in region $\rho$ traversed by region route $\phi$ . This weighting steers the optimizer toward flow configurations directly observable by sensors.
Optimization Objective: The method solves a weighted least-squares problem to find non-negative route flows ( $x_t$ $x_{t}$ ) that minimize the discrepancy between predicted regional counts ( $\hat{y}_t = C^\top x_t$ $\overset{y}{^}_{t} = C^{⊤} x_{t}$ ) and observed counts ( $y_t$ $y_{t}$ ). The objective function $L(x_t)$ $L (x_{t})$ combines two terms:
- Weighted Regional Loss ( $L_{reg}$ ): Penalizes per-region discrepancies, weighted inversely by observed counts to prevent high-volume regions from dominating the gradient.
- Aggregated Loss ( $L_{aggr}$ ): Penalizes the relative deviation between total predicted and total observed traffic volumes to ensure global consistency.
Solver: The problem is solved using the Adam gradient-based optimizer with gradient clipping. Non-negativity constraints are enforced via a softplus transformation.

2. Edge-Level Assignment (Trajectory Refinement)

Once regional flows are determined, the method refines each regional route into concrete edge-level trajectories.

Route Generation: For a regional route, representative "via" edges are selected within each region. Shortest paths are computed between consecutive via edges to form candidate edge-level routes.
Scoring and Selection: Candidate routes are scored based on two criteria:
1. Volume Coverage: The fraction of regional traffic captured by sensors along the edge route.
2. Temporal Shape Similarity: The similarity between the route's temporal demand profile and the aggregated profile of the parent regional route (measured via $L_1$ -normalized inner product).
Sampling: Routes are ranked by a combined score and converted into a probability distribution via softmax. Vehicles assigned to the regional route are then distributed across candidate edge routes by sampling from this distribution.

Key Contributions

Simulation-Free Framework: The method estimates traffic models directly from raw sensor data without requiring feedback from a traffic simulator, eliminating the computational bottleneck of iterative simulation.
No Prior OD Matrix: Unlike many state-of-the-art methods, it does not require a prior OD matrix or GPS trajectory data; it infers flows based on plausible routes on the road network graph.
Hierarchical Decomposition: By abstracting the calibration problem to spatial aggregates first, the method ensures regional concordance by construction while simplifying the optimization landscape.
Scalability and Efficiency: The problem dimensionality depends on the number of region-routes rather than network links, and the fully differentiable objective enables the use of efficient gradient-based optimizers.

Experimental Results

The method was evaluated on the Brussels road network using real traffic data from 369 sensors (March 1, 2024) and validated against synthetic scenarios.

Performance vs. Baselines:
- Proposed Method: Achieved a Mean Absolute Error (MAE) of 44.16 and RMSE of 51.80.
- Simulation-Based Baseline: Achieved MAE of 67.02 and RMSE of 80.18.
- RouteSampler (SUMO tool): Achieved MAE of 192.45 and RMSE of 237.24.
Computational Cost:
- The proposed method required approximately 23 seconds to evaluate region-level allocation over 24 hours.
- The simulation-based baseline required an average of 69 minutes to converge.
- The proposed method is approximately 180 times faster than the simulation-based baseline.
Traffic Profile Reconstruction: The method successfully reproduced the daily traffic profile (including peak structures) and the global transition matrix structure. While some edge-level errors occurred in high-density central districts due to the under-determined nature of the problem, the regional-level fit was robust.
Synthetic Validation: In synthetic scenarios with known ground truth, the method recovered the global routing structure and transition probabilities with high fidelity, though it occasionally under-diversified outflows in regions with sparse sensor coverage.

Significance and Claims

The paper claims that the proposed approach offers a practical solution for urban traffic calibration where computational resources are limited or real-time updates are necessary. By operating through a hierarchical regional structure, the method abstracts the calibration problem to a simpler approximation while ensuring regional concordance.

The authors emphasize that the method recovers a plausible and regionally consistent traffic model, rather than the unique "true" model, acknowledging the inherent under-determination of the inverse problem. The significance lies in providing a computationally efficient alternative to simulation-based calibration that does not rely on prior demand matrices or dense trajectory data, making it suitable for city-level networks with sparse sensor coverage.

Limitations Acknowledged

The paper explicitly notes three limitations:

Under-determination: Without additional data (e.g., probe trajectories), multiple flow configurations can be consistent with the data; the method recovers a plausible solution, not necessarily the ground truth.
Route Choice Heuristics: Edge-level refinement relies on shortest-route heuristics, which may not capture realistic route choices under congestion.
Sensitivity to Iterations: Excessive optimization iterations can lead to overfitting, where flows concentrate on sensor-covered routes, degrading accuracy in unobserved corridors.