travel time estimation

This study addresses the challenge of achieving both accuracy and engineering practicality in travel time prediction under low-congestion conditions. The authors propose a lightweight modeling framework that integrates open-source geographic information with sparse traffic operational features—such as traffic signals, stop signs, and turn types—to generate shortest travel-time paths using Dijkstra’s algorithm, followed by path deviation correction via random forest regression. Notably, the method operates without real-time congestion data and delivers highly accurate and efficient point-to-point travel time estimates at an urban scale. Experimental results demonstrate that the model significantly outperforms baseline approaches across multiple metrics—including MAE, MAPE, and MSE—exhibits negligible mean bias, and shows strong generalization and stability, as confirmed by k-fold cross-validation. It is thus well-suited for applications such as route planning and accessibility analysis in low-traffic scenarios.

Arc travel time estimation and path choice model parameter estimation in road transportation networks are inherently interdependent, yet traditionally addressed separately, leading to biased estimates due to ignored coupling. Method: This paper proposes the first joint maximum likelihood estimation framework, built upon a differentiable path choice model that integrates random utility theory with numerical optimization. It supports unified modeling of multi-granularity, noisy, and partially observed path data—including real-world GPS trajectories (e.g., NYC taxi data). Contribution/Results: Our approach enables end-to-end joint estimation of travel times and path choice parameters—the first such method—effectively mitigating bias inherent in sequential estimation. Experiments demonstrate significant improvements in path choice parameter accuracy and superior arc travel time estimation performance compared to link-only baselines. This framework provides a more reliable foundation for both strategic and tactical transportation network planning.

This paper addresses the ill-posedness arising from sparse observational data in simulation-driven origin–destination (OD) demand calibration. We propose a prior-free joint calibration method that directly incorporates sparse link count observations into the simulation-based optimization objective as a data-driven ℓ₂ regularization term, simultaneously constraining both link flow distribution and path travel times. Our approach establishes a tightly coupled simulation–optimization framework integrating dynamic traffic assignment (DTA) with gradient approximation algorithms to enable end-to-end calibration. Evaluated on multiple congestion scenarios within the Seattle highway network, the method achieves a 37% improvement in OD demand recovery accuracy and reduces link flow fitting error by 52%, significantly outperforming conventional prior-dependent OD calibration approaches.

Existing traffic modeling approaches—such as sensor-based forecasting, reinforcement learning, classical optimization, and demand modeling—are fragmented and lack a unified framework supporting cross-module gradient optimization, multi-dimensional and multi-modal joint analysis, and large-scale deployment. This paper proposes Flow-Through Tensors (FTT), a tensor-based unified computational graph architecture that jointly models traffic flow, route choice, and travel time. Its key contributions are: (1) a differentiable unified mathematical structure enabling end-to-end gradient propagation; (2) fine-grained spatiotemporal and demographic joint analysis; and (3) efficient tensor decomposition ensuring real-time computation on million-node networks. Experiments demonstrate that FTT significantly improves collaborative optimization efficiency and control responsiveness while strictly satisfying physical constraints.

This study addresses the high uncertainty in travel times following subway disruptions, caused by irregular train operations during service recovery. The authors propose a Bayesian spatiotemporal modeling framework that decomposes travel time into a baseline component and a delay component. Innovatively, the model jointly captures inter-train dependencies through a moving average error structure, accounts for headway imbalance, and characterizes the asymmetric heavy-tailed nature of travel times during recovery periods using skew-normal and skew-t distributions. Evaluated on high-resolution track occupancy data from the Montreal metro system, the approach outperforms existing baselines in both point prediction accuracy and uncertainty quantification, with the skew-t variant demonstrating particularly robust performance for long-distance trips.

This study addresses the critical need for high-resolution, vehicle-class-specific origin–destination (OD) data in urban traffic simulation. While mobile signaling data offer broad spatial coverage, they lack mode differentiation and suffer from representativeness bias; conversely, toll station counts provide accurate yet sparse observations. To bridge this gap, the authors propose a machine learning framework that fuses sparse, precise toll counts with large-scale mobile signaling data by modeling spatiotemporal features and incorporating route assignment logic to learn their nonlinear mapping. This approach enables, for the first time, the calibration of aggregated mobility data using minimal sensor inputs to generate hourly, vehicle-class-disaggregated macroscopic OD matrices. Validated in a bus depot expansion case study in Trondheim, Norway, the method successfully produced length-based vehicle OD estimates, significantly enhancing the accuracy and scalability of traffic flow inference—particularly in data-scarce environments.

This study addresses the challenge of estimating time-varying traffic flows under sparse vehicle count data by proposing an efficient reconstruction method that eliminates the need for traffic simulation. The approach partitions the study area into subregions, constructs a feasible path set, and introduces a sensor coverage contribution matrix to guide a weighted least squares optimization for rational path flow allocation. By further integrating temporal and volumetric characteristics of regional counts, the method generates link-level trajectories. It substantially reduces computational overhead while improving estimation accuracy, successfully reproducing daily traffic patterns in real-world experiments on the Brussels road network and outperforming existing baseline methods.

This study addresses the limited generalizability of existing traffic prediction methods across diverse route choice scenarios, which fail to accurately capture network-wide travel time variations arising from differing path allocations under identical travel demand. To overcome this limitation, the paper proposes a Generalized Travel Time Predictor (GenTTP)—the first framework capable of generalizing travel time predictions across a wide spectrum of routing strategies. GenTTP leverages graph neural networks to jointly model spatiotemporal traffic dynamics and microscopic route choice behavior, effectively capturing complex interactions within the road network. Experimental results demonstrate that GenTTP significantly improves the accuracy of both travel time and traffic flow predictions under various path assignment scenarios, exhibiting strong generalization capabilities even for atypical and dynamically evolving travel behaviors.

Existing approaches struggle to generate individual daily activity schedules that are both realistic and temporally consistent with reported travel survey data. This work proposes an iterative optimization framework that integrates dynamic programming, travel time simulation, and activity location assignment algorithms to progressively refine synthetic activity schedules while preserving privacy. By iteratively adjusting activity timing and sequencing, the method aligns simulated travel times with empirical distributions derived from travel surveys. Experimental results demonstrate that the proposed approach reduces the discrepancy between simulated and observed travel times by 52.2% compared to initial schedules, substantially improving both temporal consistency and behavioral realism in the generated activity plans.