Human lane-changing behaviour is highly variable, making accurate intention prediction and socially adaptive decision-making challenging for intelligent vehicles. This capability is essential for safe and effective human–machine collaboration. However, real-world driving data are often degraded by communication loss between vehicles and roadside units, as well as sensor noise and uncertainty. To address these challenges, this study proposes CMT-GAT, a robust self-supervised contrastive learning framework for lane-change intention prediction in urban expressway weaving areas. The model employs a dual-channel trend attention mechanism to process complete and incomplete data in parallel, while a graph attention network captures interactions among surrounding vehicles and environmental context. Experiments demonstrate that CMT-GAT consistently achieves the highest prediction accuracy across all prediction horizons and missing-data rates, particularly under severe missing-data conditions.
The standard negative binomial (NB) and negative binomial-Lindley (NBL) models may not completely capture variations associated with time and space. To address these limitations, this study derives an enhanced version of the NBL model that simultaneously incorporates spatiotemporal random parameters to account for three key factors: temporal variations, spatial variations, and datasets with a large amount of zero observations. The proposed framework extends the standard NBL model by allowing both model coefficients and Lindley parameters to vary over time and across space together. This approach captures the heterogeneity in crash data caused by variations in spatiotemporal factors associated with traffic patterns, environmental conditions, or geographic differences, while also addressing the high proportion of zero observations typically found in such datasets. The results demonstrate that the proposed model reasonably recovers the true parameters. Findings indicate that the proposed model outperforms those that solely account for temporal or spatial variations.
Air cargo transportation via the unused belly capacity of passenger flights can improve transportation efficiency and reduce operational costs. To realise these potential benefits, we propose a framework to efficiently manage unused capacity on passenger flights, accommodating uncertainties in flight capacity and cargo demand. A two-stage stochastic program is proposed for the studied problem, with the first stage introducing a mid-term capacity reservation scheme while the second stage focuses on profitable short-term cargo-to-flight allocations. We solve the two-stage problem with a Progressive Hedging decomposition and estimate the expected cargo-to-flight assignment cost via Sample Average Approximation. To construct high-quality allocations, several heuristic algorithms are tested within the cargo assignment stage. Using Hong Kong International Airport as the origin and 8 closely connected destinations, we collect real flight schedules and simulate demand data. Extensive experiments on medium-/large-scale instances show that the model converges reliably and the solution approach scales effectively. In addition, a sensitivity analysis is conducted, which highlights the importance of efficiently managing capacity in passenger flights. Relative to a benchmark that reserves all available belly capacity, the two-stage framework reduces operating costs even when the unit revenue for post-reservation capacity is unchanged. Across routes, costs fall by roughly 7–16% in most cases and by about 30% on the best-performing market. When unit revenue improves and residual capacity is ample, savings increase substantially, rising to roughly 26–57% and, in some markets, exceeding 50–80%, with one route even generating a surplus (negative cost), indicating revenue fully covers expenses.
CHANAkYA is a comprehensive game-theoretical framework for pricing and tradable credit charging in complex transportation systems. It models strategic interactions among public authorities and private mobility providers, while incorporating multimodal travel options and heterogeneous traveller demographics. It supports link- and path-based tolls, alongside equivalent tradable credit schemes and allows demographic-differentiation. Additionally, the framework addresses the non-uniqueness of user-equilibrium path flows by incorporating player-specific perspectives. Five policy-driven case studies inspired by Leuven demonstrate its applications. Results indicate an optimal cordon toll of approximately €1 and a near-zero optimal bus fare, producing substantial efficiency gains over the base scenario. Age-differentiated tolling offers only modest efficiency improvements while raising equity concerns. Tradable credit schemes replicate the efficiency gains of pricing. Further cases reveal efficiency losses from privatizing public transport and show how the government’s perspective influences optimal policies. Overall, CHANAkYA supports nuanced, effective, and equitable transport planning across complex strategic and multimodal settings.
Designing road pricing in multi-decision-maker scenarios is computationally demanding, particularly when using full-scale traffic models. This paper presents a novel metamodel-based equilibration (MBE) scheme that leverages a fast game-theoretical metamodel to compute optimal/equilibrium tolls for an underlying setup of full-scale traffic models. MBE follows an iterative calibration–equilibration cycle: the metamodel is calibrated using full-scale outputs, game-theoretical equilibrium is computed by the metamodel, and the resulting tolls are passed back to the full-scale model, until a desirable outcome is identified. Unlike conventional surrogate approaches that approximate a single objective function, the metamodel in MBE approximates the entire transport system, thereby shifting the computational burden of toll equilibration to the metamodel, and enabling efficient analysis in multi-decision-maker settings. Applied to a real-world network with variable demand, MBE outperforms Bayesian Optimization for single-decision-maker problems and, for the first time, computes Nash and Stackelberg equilibria for multi-decision-maker problems in a full-scale traffic model.
Accurately predicting pedestrian crossing intentions is essential for safe autonomous driving, yet generalisation across diverse environments remains challenging because pedestrian behaviour varies with cultural norms, traffic regulations, and individual decision-making. This paper proposes a behavioural domain-adaptive spatio-temporal graph convolutional network (BDAST-GCN) for pedestrian intention prediction. The framework includes two components. First, a behavioural style representation module uses an enhanced Trans-TimeGAN with dimensionality reduction and clustering to identify latent behavioural domains from dynamic pedestrian features. Second, a domain-adaptive prediction module employs a spatio-temporal graph convolutional network with distribution matching loss to align feature representations across domains while incorporating vehicle speed and contextual information. Experiments on the JAAD and PIE datasets demonstrate that BDAST-GCN effectively predicts pedestrian intentions and improves generalisation across heterogeneous traffic environments. These results indicate that the proposed framework provides a robust and adaptive solution for real-world pedestrian intention prediction in autonomous driving.
A Neural-Network-embedded Generalised Ordered Response Probit (NNGORP) model captures nonlinear effects on latent propensity for ordered activity frequency outcomes. Jointly learning latent propensity and thresholds, it embeds neural representations into a probabilistic latent-index structure for counts, preserving a coherent discrete distribution over non-negative integers while capturing complex nonlinearities. Simulations verify accurate parameter and marginal effect recovery from a benchmark GORP. It models discretionary activity frequency using 60,309 tours from the 2019 Shanghai Household Travel Survey including sociodemographic, household, and built-environment variables. Results reveal nonlinearities and threshold responses for age, household employment, young children, bicycle ownership, and neighbourhood density. Structurally derived from the latent-index mechanism, partial dependence patterns reflect genuine behavioural processes rather than post-hoc approximations. Behavioural interpretation confirms structurally derived patterns provide more coherent, generalisable insights than SHAP from conventional ML. These findings deepen understanding of discretionary activity generation in high-density urban settings, informing travel demand management and planning.
This study develops a fully flexible demand responsive transport (DRT) system that processes incoming requests immediately without advance reservations, fixed routes or stops, or prior demand aggregation. An insertion heuristic evaluates feasible vehicle assignments and pick-up and drop-off insertion positions subject to capacity and time-window constraints and pick-up-before-drop-off precedence, selecting the alternative with the smallest increase in vehicle travel time. The system was evaluated across 100 simulation runs on the Sioux Falls network with 80 requests per hour and four vehicles. Mean computation time per run was 2.14 s, and the mean request acceptance rate was 58.46%. Accepted passengers experienced mean waiting and detour times of 6.17 and 4.20 min, respectively. Increasing the fleet to eight vehicles raised the acceptance rate to 93.64% under the same demand. These findings demonstrate the computational feasibility of the proposed framework for real-time, fully flexible DRT operations.
This study develops a fractional-order approximation framework for static traffic equilibrium under stochastic network-element uncertainty. Random OD demand, link flow, link capacity, or travel time is propagated through nonlinear impedance functions, and non-integer BPR exponents require fractional moments of random variables. To improve upon conventional Taylor expansion, the proposed method combines fractional Taylor expansion with Padé rational approximation and introduces unified control of truncation, Padé, and re-expansion errors. Practical procedures for parameter selection and iterative implementation are also provided. Numerical experiments on the Nguyen and Dupuis and Sioux Falls networks examine stochastic link-flow and link-capacity settings under Gaussian mixture, inverse Gaussian, lognormal, Gamma, Weibull, and Pareto distributions. Results show lower and more stable travel-time approximation errors, more reliable equilibrium path flows and travel times, and higher computational efficiency than Taylor expansion and Monte Carlo sampling, particularly under strict convergence criteria and larger path sets.
In the study of smart rail transit systems, demand prediction, operational optimisation, and resilient network planning are essential, as urban rail increasingly operates within complex multimodal and uncertain environments. This special issue highlights the role of artificial intelligence and large-scale data analytics in advancing smart rail transit, and examines the topic through forecasting, control, disruption management, multimodal coordination, and decision optimisation. The works comprising this special issue can thus provide ideas for future studies on intelligent rail operations and serve as a long-term reference for data-driven smart transportation systems.
Amid the rapid expansion of global chemical trade, liquid chemical ports currently suffer from low automation and reliance on manual experience. To address the high coupling of berths and pipelines, this study establishes a mixed-integer programming model for collaborative berth-pipeline allocation, integrating three operation modes (direct-connected pipelines, berth-connect pipelines, and manifold-connect pipelines) to minimise total vessel port time. A tailored memetic algorithm with adaptive local search (MA-ALS) is designed, incorporating problem-specific operators to enhance optimisation capability. Numerical experiments demonstrate that our proposed method outperforms other baselines. Furthermore, the study advises prioritising berth upgrade redundancy and identifying marginal benefit inflection points to optimise throughput capacity scientifically.
The superimposed impacts of bridge and fog on car-following risk remain unclear. The study proposed a risk-field-guided quantitative framework integrating risk assessment and prevention. We designed the ordinary road, ordinary bridge, and foggy bridge to conduct driving simulation experiments. Risk field force models were developed to quantify car-following risk. A targeted prevention strategy was designed and validated for the foggy bridge. Results reveal that: (1) Bridges exacerbate the heterogeneity in speed and impairs drivers′ cognition. Fog destabilises the car-following process and deteriorates behaviour performances on bridges; (2) On bridges, lower car-following risk reflects compensatory driving behaviours, rather than lower risk due to bridge-specific road conditions. On foggy bridges, drivers exhibit excessive following tendencies with inadequate behavioural compensation, resulting in the highest risk; (3) The prevention strategy effectively alleviates car-following risk on foggy bridges. The study offers a theoretical foundation and application guidance for bridge safety management.
Real-time and accurate traffic state estimation is crucial for intelligent freeway management. To fully utilize existing resources and achieve high-resolution estimation, this study proposes a multi-source heterogeneous fixed sensor data fusion framework. First, interpolation is adopted to unify reporting frequency to one minute, improving time alignment and data continuity. Then, four deep learning models, including Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Stacked Auto-Encoders (SAE), and Transformer, are employed for traffic state estimation. To achieve optimal performance, twelve input step sizes and five interpolation methods are considered. The framework is comprehensively evaluated using real-world freeway data. Results show that each model has a distinct optimal configuration, and both interpolation and data fusion contribute to improving estimation performance. Multi-dimensional analysis further confirms the model’s robustness under varying traffic states, freeway section types, weather conditions, sensor failures, and cross-segment transfer test, enabling effective support for timely traffic management decisions.
For effective work zone safety, it is crucial to link safety with flow dynamics. This study proposes a Extreme Value Theory (EVT) framework to quantify crash risk from rear-end conflicts and assess work zone safety using macroscopic flow dynamics. Traffic conflicts were extracted from videos from an urban expressway, covering work zone and baseline conditions. The extremes were modelled using the Block Maxima approach with several covariates incorporated into the parameters of the Generalized Extreme Value distribution. A novel metric, Relative Work Zone Impact (RWZ), was introduced to quantify the risk elevation due to work zone presence. RWZ values were mapped onto traffic flow relationships to construct a Safety-Aware Fundamental Diagram (SAFD). The results indicate that work zones significantly compromise safety during intermediately congested flow conditions, where traffic instability prevails. The RWZ-SAFD identifies high-risk traffic states and supports safety measures like variable speed limits and adaptive work zone scheduling.
Understanding autonomous vehicle (AV) crash mechanisms is critical for safety, yet linear assumptions may overlook complex risk patterns. An interpretable machine learning framework was employed to examine the non-linear effects of various factors on AV crash severities based on 673 crash reports from the California Department of Motor Vehicles (CA DMV) augmented with Point of Interest (POI) data. Specifically, quarter, time of day, hospital density, AV movement, traffic signals nearest distance, weekday, driving mode, parking density, vulnerable road users (VRUs), and parking nearest distance show important non-linear impacts. Interaction and PDP analyses further reveal conditional and marginal effects across factors. The contribution of this study lies in how built environment attributes non-linearly affect AV crash severity, offering an interpretable framework for safety management. Findings inform context-sensitive algorithm adjustments and infrastructure planning. Future research could integrate sensor and real-time data to shift from retrospective analysis to proactive, predictive safety systems.
Time-use decisions are a critical input for quantifying travel demand in activity-based models. These decisions are typically modelled using econometric methods grounded in utility maximisation theory - multiple discrete-continuous extreme value (MDCEV) models in particular. Since time-use decisions are quite complex and may not necessarily be underpinned by a utility-maximisation framework, we propose to use a data-driven machine learning technique (Recurrent Neural Networks, RNN) to model time-use. We compare and contrast time-use models developed using a long short-term memory (LSTM), a type of RNN model, with an MDCEV model based on the predictive performance and marginal effects. Using time-use data collected during the COVID-19 pandemic in the UK, we observe that there are no significant differences in prediction accuracy between the two approaches, both at the aggregate (sample) and disaggregate (individual) levels. Thus, we find no evidence that data-driven methods outperform traditional econometric models in predicting time-use behaviour. We also observe that the marginal effects derived from both models are broadly similar. However, in the absence of benchmark or ground-truth values, it is not possible for us to conclude the output of which model is closer to reality. This limits the depth of the comparison. Overall, our findings suggest that RNN models can serve as viable alternatives to MDCEV models for modelling time-use with the added capability of generating activity schedules. In contrast, MDCEV models offer greater transparency about the assumptions and provide more interpretable, policy-relevant outputs.
With the continuous development of autonomous driving technology, accurate trajectory planning plays a crucial role in improving the safety decision-making ability of autonomous vehicles. This paper proposes a joint planning method combining multimodal trajectory prediction and spatiotemporal driving corridors, addressing the limitations of existing methods in using historical trajectory data and modelling multi-vehicle interactions. The method uses a Gaussian mixture model (GMM) to predict surrounding vehicles' future trajectories, evaluates collision risks based on asymmetric interactions, and combines these predictions with spatiotemporal driving corridors to generate optimal paths while avoiding collision risks. The system initially generates a trajectory using dynamic programming (DP) and optimises it via quadratic programming (QP) to ensure smooth, safe, and efficient paths. Compared to the RS + QP algorithm in high-density traffic, the proposed algorithm improves vehicle speed by 29.37% and reduces longitudinal acceleration variance by 54.9%, demonstrating better efficiency and comfort.
