Abstract
The estimation of origin-destination (OD) matrices is a crucial aspect of Intelligent Transport Systems (ITS). It involves adjusting an initial OD matrix by regressing the current observations like traffic counts of road sections (e.g., using least squares). However, the OD estimation problem lacks sufficient constraints and is mathematically underdetermined. To alleviate this problem, some researchers incorporate a prior OD matrix as a target in the regression to provide more structural constraints. However, this approach is highly dependent on the existing prior matrix, which may be outdated. Others add structural constraints through sensor data, such as vehicle trajectory and speed, which can reflect more current structural constraints in real-time. Our proposed method integrates deep learning and numerical optimization algorithms to infer matrix structure and guide numerical optimization. This approach combines the advantages of both deep learning and numerical optimization algorithms. The neural network(NN) learns to infer structural constraints from probe traffic flows, eliminating dependence on prior information and providing real-time performance. Additionally, due to the generalization capability of NN, this method is economical in engineering. We conducted tests to demonstrate the good generalization performance of our method on a large-scale synthetic dataset. Subsequently, we verified the stability of our method on real traffic data. Our experiments provided confirmation of the benefits of combining NN and numerical optimization.
Introduction
Managing urban traffic effectively requires an accurate Origin-Destination (OD) matrix—a fundamental tool for planners that quantifies trip volumes between specific start and end points. However, direct measurement of these matrices is often impossible. Instead, planners rely on traffic counts from road sensors. The core challenge is that OD estimation is mathematically underdetermined: the number of possible OD pairs vastly exceeds the number of available sensors, meaning many different trip patterns can produce identical sensor readings. Traditional methods often rely on outdated historical surveys or expensive real-time tracking hardware. This paper proposes a hybrid framework that leverages deep learning to infer structural constraints, effectively "filling in the blanks" without relying on stale prior data.
Methodology: A Hybrid Framework
The proposed method utilizes a two-stage pipeline that merges the pattern-recognition capabilities of deep learning with the precision of classical numerical optimization.
- The Distribution Learner: At the heart of the framework is a Transformer-based architecture designed to capture spatial dependencies between city zones. This model takes the road network topology and current traffic count distributions as inputs. Crucially, instead of predicting raw trip volumes—which can vary by several orders of magnitude—the model predicts the probability distribution of trip production and attraction. This normalization makes the learning process more stable and generalizable across different urban scales.
- Numerical Optimization: These inferred distributions serve as structural constraints in a bi-level optimization framework. The upper level adjusts the OD matrix to minimize the discrepancy between observed and simulated traffic counts, while the lower level handles traffic assignment. The authors employ Jensen-Shannon Divergence (JSD) for training the neural network and reverse Kullback-Leibler Divergence (KLD) during optimization to ensure the final matrix adheres to the most significant trip patterns identified by the learner.
Datasets and Evaluation
The framework was rigorously tested on two datasets representing different scales and data types:
- Cologne, Germany (Synthetic): A massive dataset covering 400 $km^2$ with over 71,000 road sections. Ground-truth patterns were generated using the SUMO (Simulation of Urban MObility) simulator.
- Bologna, Italy (Real): The "Ringway Dataset," providing real-world traffic data from 15 OD zones and 55 road sections, used to validate the model's stability against noisy, real-world observations.
Key Results and Discussion
The hybrid approach demonstrated significant performance improvements over traditional least-squares methods:
- Accuracy and Scalability: In a 15x15 grid in Cologne, the method improved the Root Mean Square Error (RMSN) by 17.58%. As the problem scale increased to a 50x50 grid, the improvement jumped to 42.12% in RMSN and 38.37% in structural similarity ($\rho$).
- Robustness: The model maintained high performance even when the initial "guess" (prior matrix) was heavily corrupted with noise, indicating that the system relies on learned structural intelligence rather than historical averages.
Conclusion
By combining the global structural inference of Transformers with the local fine-tuning of numerical optimization, this research provides a scalable and economical solution for real-time OD matrix estimation. While the method's success depends on the accuracy of the Distribution Learner, its ability to generalize across large-scale networks without outdated priors marks a significant step forward for Intelligent Transport Systems. Future work will focus on integrating real GPS trajectory data to further enhance routing accuracy.
Summary updated