Q-Net: Transferable Queue Length Estimation via Kalman-based Neural Networks

Ting Gao, Elvin Isufi, Winnie Daamen, Erik-Sander Smits, Serge P. HoogendoornarXiv 2025 · DOI · arXiv

Summary

Q-Net is a hybrid framework that combines the physical interpretability of Kalman filters with the flexible learning power of Gated Recurrent Units (GRUs) to estimate traffic queue lengths. By fusing high-frequency vehicle counts with low-frequency speed data, it achieves a ~24% improvement in accuracy over state-of-the-art baselines and remains spatially transferable across different road geometries without retraining.

The Challenge of Urban Traffic Observability

Estimating the length of vehicle queues at signalized intersections is a cornerstone of intelligent transportation systems. However, in urban environments, this task is complicated by partial observability. Data typically comes from two disparate sources: loop detectors, which provide precise vehicle counts at specific points, and aggregated Floating Car Data (aFCD), which offers average speeds from GPS-equipped vehicles but at lower temporal resolutions (e.g., 1-minute intervals).

A significant challenge is the violation of traffic conservation laws due to unmonitored "leakage"—vehicles entering or exiting via side streets or parking lots. Traditional filtering methods often fail to account for this non-linear noise and the drift caused by these unobserved flows.

Methodology

The proposed Q-Net architecture treats queue length as a hidden state within a state-space model, leveraging three core pillars:

  • AI-Augmented Kalman Filtering (KalmanNet): Standard Kalman filters require analytical calculation of the Kalman Gain ($K_t$), which depends on often-unknown noise statistics. Q-Net employs a neural network (specifically GRUs) to learn this gain directly from data, allowing the model to adapt to complex, time-varying noise patterns.
  • Robust Control Input: To mitigate the effects of "leaky" traffic flows, the model derives a "queue change" input from loop detectors. This input is refined using an affine transformation and a Fourier band-pass filter to eliminate low-frequency drift and high-frequency sensor noise.
  • Spatial Transferability: Unlike many deep learning models that are tied to specific road geometries, Q-Net uses a "local grouping" strategy. By processing speed measurements in fixed-size neighborhoods (previous, current, and next segments), the input dimension remains constant, enabling zero-shot transfer to new road sections.
Figure 1: Schematic of the study section showing inflow/outflow detectors and aFCD segments used for queue length estimation.

Experimental Setup and Evaluation

The framework was evaluated using real-world data from Rotterdam, the Netherlands, collected in November 2023. The dataset included 10-second loop detector counts and 60-second aFCD speed data. Ground truth was established using high-accuracy radar sensors providing measurements every 10 seconds across three major road sections (N1-IN, N3-IN, and N3-OUT).

Key Results

Q-Net demonstrated superior performance across all test scenarios, reducing the Root Mean Square Error (RMSE) by 24.05% and Mean Absolute Error (MAE) by 14.44% compared to the strongest baseline (Tree-RPROP).

  • Peak Performance: The model was particularly effective during morning peak hours, achieving a 29% reduction in RMSE where traffic patterns are most volatile.
  • Zero-Shot Transfer: A model trained on one road section (N1-IN) was successfully applied to others (N3-IN/OUT) without retraining, accurately capturing queue dynamics despite different geometries.
  • Efficiency: With only 952 trainable parameters, Q-Net is highly efficient and suitable for real-time edge deployment on modest hardware.
Figure 4: Comparison between ground truth queue length and the reconstructed queue length using the Q-Net framework.

Conclusion

Q-Net bridges the gap between physics-based modeling and deep learning. By embedding a neural network within a Kalman filter framework, it provides a robust, interpretable, and highly transferable solution for real-time urban traffic management. Future work will explore the integration of traffic signal timings (SPaT data) and testing across diverse international urban environments to further validate its generalizability.

Summary updated

BibTeX
@misc{gao2025q,
  author = {Ting Gao and Elvin Isufi and Winnie Daamen and Erik-Sander Smits and Serge P. Hoogendoorn},
  title = {{Q-Net: Transferable Queue Length Estimation via Kalman-based Neural Networks}},
  year = {2025},
  eprint = {2509.24725},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.24725},
}
Figure