Vehicle re-identification (Re-ID) plays a vital role in vehicle-infrastructure cooperation systems, enabling continuous trajectory association across non-overlapping roadside camera views and providing essential perception redundancy for autonomous driving. However, in complex urban environments, frequent occlusions caused by surrounding vehicles, roadside structures, and pedestrians severely degrade Re-ID performance. Existing external model–based solutions often depend on auxiliary segmentation or detection networks, which increase computational overhead and hinder deployment in resource-constrained roadside units. In contrast, model-free approaches based on image enhancement provide limited gains when addressing structural occlusions, while multi-image feature fusion methods require additional images for joint training, substantially increasing training complexity. To address these challenges, this study proposes a novel dual-branch feature reconstruction model for the Re-ID (DFRR) framework, which integrates a dual-branch image feature encoder for discriminative feature extraction and an image reconstruction module for occluded image restoration under diverse and enhanced occlusion scenarios. Without relying on external models or neighboring image information, DFRR effectively recovers structural details in occluded regions, enabling robust single-image Re-ID in real-world traffic conditions. Extensive experiments on the VeRi-776 and DAIR-V2XReid datasets demonstrate the effectiveness of the proposed framework, achieving a higher mean average precision and delivering competitive performance compared with benchmark methods, particularly under severe occlusion conditions.
Accurate traffic flow information is crucial for effective operation of adaptive signal control systems. However, conventional approaches predominantly rely on deterministic point estimation methods, which fail to capture the inherent uncertainties in real-world traffic environments, thereby limiting the robustness of signal control decisions. This study introduces an uncertainty-aware prediction-control framework that leverages a probabilistic Bayesian approach to predict short-term traffic flows and adjust signal control plans adaptively. Specifically, a Dirichlet Gaussian Mixture Model (DGMM) is formulated to characterize the joint probabilistic distribution of traffic variables at signalized intersections, integrating local traffic flows, signal settings, and upstream inflows across consecutive cycles. In contrast to deterministic methods, this framework enables both interpretable traffic pattern identification and probabilistic short-term flow forecasting. These probabilistic forecasts are subsequently integrated into a two-stage stochastic programming algorithm devised to determine adaptive signal timings under demand uncertainty. The prediction performance of the proposed DGMM is evaluated against statistical and deep learning benchmarks using real-world data collected from Tung Chung, Hong Kong. Simulation results demonstrate that the uncertainty-aware control strategy can reduce average vehicle delays by up to 22.8 % during high-variability peak hours compared to deterministic approaches. Overall, the proposed framework offers a practical and robust solution for uncertainty-aware adaptive signal control in complex urban networks.
Accurate fuel consumption prediction is often limited by data scarcity, especially as current state-of-the-art learning-based approaches are typically data-hungry. In practice, historical records are unevenly distributed across the fleet. While a few well-equipped vessels provide dense, high-frequency data, most vessels have only sparse records due to short observation periods, manual log collection, or maintenance-related data gaps. Consequently, training robust models for vessels with limited historical data is challenging. To address this challenge, we propose F2V-Former, a novel fleet-to-vessel transfer learning framework that leverages abundant data collected across the fleet to pre-train a unified model before fine-tuning it for target vessels. To improve adaptability to heterogeneous vessels with different capacities, F2V-Former integrates static vessel attributes with dynamic operational sequences via static-guided conditional layer normalization and cross-attention. For parameter-efficient vessel-specific adaptation in few-shot settings, we adopt low-rank adaptation (LoRA) to fine-tune the network. Experiments on a real-world container shipping fleet show that F2V-Former achieves the best performance in zero-shot and in most few-shot settings. Notably, in zero-shot experiments, F2V-Former achieves a MAPE of 14% without vessel-specific fine-tuning, demonstrating its reliable generalization among the new vessels in the fleet. Further experiments, including error evolution analysis, sensitivity analysis, and ablation studies, are conducted to validate the effectiveness of the proposed framework. These results demonstrate that F2V-Former achieves high prediction accuracy and reliable cross-vessel generalization in data-scarce scenarios.
Accurately recognizing diverse driving behaviors, with particular emphasis on safety–critical merging maneuvers, is crucial for improving road safety and traffic efficiency. Reliable predictions allow automated and connected vehicles to anticipate potential conflicts, choose safe gaps, and keep traffic flowing smoothly even in busy conditions. However, most existing methods treat all drivers the same and ignore factors such as mental workload and individual driving style, which limits their realism and accuracy. To address these limitations, we propose CAP-Drive, a Complexity-Aware Personalized Driving framework that combines real-time estimation of a behaviorally grounded operational demand indicator, a Controller Area Network (CAN)-bus proxy of driver workload, with personalized driving style profiles, analyzed through a transformer-based deep learning model. CAP-Drive jointly learns to recognize behavior, estimate operational demand, and predict merge confidence, while translating these insights into real-time, personalized assistance. Tests using the Honda Research Institute Driving Dataset show that CAP-Drive outperforms existing methods in predicting driving behaviors and provides more reliable, well-calibrated confidence estimates. In short, CAP-Drive offers a new step toward intelligent driver-assistance systems that are personalized, adaptive, and trustworthy.
Decentralized control has been a central focus in the management of large-scale urban traffic networks. Among various strategies, the Max-Pressure algorithm has emerged as a leading method for decentralized signal control. Notably, it operates without requiring prior knowledge of traffic demand, while implicitly assuming that arrival processes are admissible within the network’s stability region. Theoretical guarantees, including throughput optimality, are established under idealized assumptions that neglect signal control constraints and are largely inherited from its origins in communication networks. This study aims to characterize network capacity by defining the stability region for each intersection under full implementation of the Max-Pressure policy, while accounting for realistic traffic signal constraints such as inter-green periods and minimum green activation times. We focus on simple two-phase intersections, for which the signal stage sequence is fixed by design and does not require explicit sequencing constraints. As a foundational step, we first analyze the Max-Pressure controller in an isolated intersection setting. Using both analytical derivations and simulation, we evaluate the impact of temporal constraints - specifically inter-green times and minimum green durations - on the controller’s ability to maintain system stability. Our results demonstrate that once these practical constraints are introduced, the classical throughput-optimality guarantees no longer hold in general, and capacity losses emerge. The stability and control activation analysis provides essential insights into the structural limitations of Max-Pressure control under practical conditions, and serves as a stepping stone for extending the stability analysis to complex urban traffic networks.
Autonomous driving has progressed rapidly in recent years; however, its performance in complex conflict zones remains a significant challenge. With the advancement of connectivity technologies, roadside infrastructure, leveraging its global perspective, can provide vehicles with enhanced information and decision support. Nevertheless, existing cooperative frameworks often overwhelm vehicles with excessive raw information while lacking concise semantic abstraction, and they insufficiently account for heterogeneous vehicle demands. To address these challenges, this paper proposes the Vehicle-Infrastructure Language-Enhanced Negotiation and Cooperation (V2I-LENC) chain for mixed-traffic conflict zones. It first constructs a semantic-level scene interpreter using a VLM fine-tuned on a self-built BEV dataset, enabling rich semantic understanding of the current traffic scenario. Then, an LLM-enabled hierarchical negotiation module is further used to generate the passing order by balancing global coordination with individual vehicle demand through structured negotiation. In addition, a discriminator module is introduced to assess the logical consistency and rationality of intermediate negotiation opinions, as well as the resulting passing order, across different stages of the negotiation process. The finalized passing order is ultimately translated into executable control commands by a low-level controller. Comparative and ablation experiments in CARLA demonstrate that V2I-LENC outperforms mainstream baselines in safety, efficiency, and satisfaction of passing demands, while maintaining stable performance across varying traffic scales and Connected Automated Vehicle (CAV) penetration rates.
We study a dynamic type-matching problem in mobility-on-demand (MoD) systems where multiple passenger segments have differentiated mobility requirements and multi-type vehicle fleets exhibit asymmetric service compatibilities. This general structure arises naturally whenever a specialized vehicle type is exclusively reserved for a high-priority demand segment and must be prioritized accordingly. Motivated by the mobility challenges of aging societies, we instantiate this structure in a senior-centric MoD context, incorporating Wheelchair Accessible Vehicles (WAVs) as a complementary fleet component dedicated to serving seniors with accessibility needs. The resulting type-matching problem features multi-type fleets, differentiated passenger segments, and asymmetric compatibilities, necessitating the development of heterogeneous yet interdependent fleet policies. We formulate the problem as a stochastic dynamic program that embeds accessibility and priority principles, and develop a Fleet-Decomposable Markov Game (Fleet-Dec MG) to model coordinated decisions among fleets with asymmetric roles. We establish a sequential fleet-based value decomposition property that preserves the long-run objective while respecting feasibility coupling between fleet policies. Building on this, we propose a Fleet-Decomposable Trust-Region Policy Optimization (FDTRPO) algorithm as a theoretical solution to the proposed Fleet-Dec MG. We prove that FDTRPO has monotonic improvement properties and converges to a Trust-Region Nash equilibrium. To enable scalable real-time deployment, we further develop a Fleet-Decomposable Proximal Policy Optimization (FDPPO) algorithm. Through city-scale experiments on real-world taxi data, we demonstrate that FDPPO converges smoothly in a high-dimensional, stochastic environment and yields deployable policies that enhance both efficiency and fairness. Compared with advanced learning benchmarks, the optimal FDPPO policy achieves a 2.6% increase in system profit and boosts the order fulfillment rate for senior passengers to over 90% with an improvement of nearly 9 percentage points for those with accessibility needs, while the levels of service of other segments are also improved. We also derive managerial insights for operators aiming to expand accessible and inclusive services for seniors and passengers with disabilities. Overall, this work advances a theoretically grounded yet practical paradigm for the dynamic type-matching problem with asymmetric compatibilities and interdependent fleet coordinations.
Airline demand forecasting for passengers and baggage is critical for resource and service planning. As airline networks move toward higher density and complexity, travel patterns vary widely across routes, and historical demand is hard to share or transfer. At the same time, frequent route launches, seasonal shifts, and schedule changes create many cold-start routes with scarce data, posing significant challenges for forecasting. Under few-shot conditions, previous forecasting methods lack sufficient accuracy and stability, and deep learning models are uninterpretable and fail to incorporate aviation business rules or quantify uncertainty. To address these challenges, this study proposes FlightLLM, an LLM-driven two-phase framework for few-shot airline demand interval forecasting that integrates business-feasibility constraints and business-related priors into the learning and adaptation process. The offline phase generates high-quality pseudo-labels using a five-stage search guided by process reward modeling, followed by curriculum-based self-training and meta-learning to build a generalized initializer. In the online phase, the model rapidly adapts to cold-start routes using minimal real-world samples. It produces joint passenger and baggage forecasts in an interpretable “point estimate ± interval half-width” format. The proposed method introduces a new way for few-shot demand forecasting by leveraging large language models to unify data-driven generalization and rule-constrained generation, where business-feasibility constraints and business-related priors are embedded into interval prediction via meta-initialization and reliability-aware adaptation. Empirical evaluations on real cold-start route data show that FlightLLM achieves a substantially more favorable coverage–width trade-off than conventional interval baselines while maintaining competitive point-forecasting accuracy, underscoring the insight that effective few-shot demand forecasting requires the complementary strengths of empirical adaptability and operational consistency.
As the “power heart” of Urban Rail Transit (URT), the Traction Power Supply System (TPSS) is designed to provide sufficient traction power for trains. However, TPSS faults often lead to a Unidirectional Power Supply Shortage (UPSS), where the total power demand for trains is restricted within a local area. To mitigate the impact of UPSS on train operations, this paper proposes a distributed train rescheduling approach. The global problem is decomposed into three sub-problems through geographical and temporal partitioning, with tailored rescheduling models developed for each. For under-supplied area, a cooperative control model is developed to optimize train control strategies while maximizing regenerative power utilization. In well-supplied areas, a two-stage Train Timetable Rescheduling (TTR) model is proposed to maximize line capacity utilization and minimize the total waiting time of passengers using various rescheduling measures. Based on the initial results of the first two sub-problems, a feedback adjustment mechanism ensures global feasibility by synchronized times and the corresponding train trajectories in the cooperative control results. Finally, the TTR model is introduced for the recovery period to restore planned train operations. Given the safety-critical nature of train operations, a Hybrid-Shield Deep Q-Networks (HSDQN) algorithm is designed by enhancing the classical DQN with multiple protection mechanisms. Specifically, a preemptive shield enforces minimum headway constraints, while a post-posed shield addresses power supply capacity constraints. In addition, transition experiences with penalty terms are stored in the replay buffer to reduce the likelihood of unsafe actions. Finally, two numeral case studies based on Beijing Metro Yizhuang Line are presented to validate the proposed approach. Results show that HSDQN outperforms three reinforcement learning algorithms (CMB-IDQN, PDQN and SDQN) by up to 12.8%, 9.0%, and 2.6%, respectively. Compared to the previous TTR approach based on maximum traction power, the proposed distributed train rescheduling approach achieves performance improvements in both line capacity utilization and passenger waiting time.
Takeover failure in conditionally automated driving is rarely caused by a single erroneous action; instead, it emerges through a causal chain in which risk is transmitted and accumulated across successive takeover process. Uncovering this mechanism is therefore essential for safe human-machine interaction. However, modeling the takeover process remains challenging because it involves multimodal and time-varying factors, lacks an explicit process-level mediator, and requires causal attribution under chain-like propagation. To address these challenges, this study proposes DBSE-SOR C2Net, an interpretable and inferable causal-chain network grounded in the cognitive framework of stimulus(S)-organism(O)-response(R). Specifically, the takeover reaction process (O) is introduced as the key mediator linking the pre-takeover states (S) to the post-takeover performance outcomes (R), and is further decomposed into nodes of the postural, perceptual and cognitive stage (PPC). To capture multimodal temporal characteristics, a multi-timescale encoding scheme is developed to jointly model long-horizon baselines and short-horizon perturbations in physiological and environmental signals. On this basis, a hierarchical dynamic Bayesian structural equation solution is constructed, with a bottom measurement layer for latent construct mapping and a top mechanism layer for dependency learning. Based on a human-in-the-loop takeover experiment (n = 41), the proposed framework effectively uncovers representative high-risk failure chains, including time-pressure-driven, cognitive-bottleneck, and multi-stage prolongation patterns. Under this framework, the adaptive takeover requests enhancement reduced PPC-stage reaction times by 0.11 s, 0.07 s, and 0.45 s, and shortened the overall takeover time by 0.31 s. Moreover, downstream PPC-stage monitoring with Minimum Risk Maneuver intervention reduced the collision rate by 75%. In general, the DBSE-SOR C2Net provides a backward reasoning and forward propagation paradigm for failure-chain diagnosis and adaptive safety intervention in automated driving.
Sufficient testing under corner cases is critical for the long-term operation of vehicle-infrastructure cooperation systems (VICS). However, existing corner case generation methods are primarily AI-driven, and VICS testing under corner cases is typically limited to simulation. In this paper, we introduce an L5 “Interactable” level to the VICS digital twin (VICS-DT) taxonomy, extending beyond the conventional L4 “Optimizable” level. We further propose an L5-level VICS testing framework, IMPACT (Interactive Mixed-digital-twin Paradigm for Advanced Cooperative vehicle-infrastructure Testing). By enabling direct interactions between human operators and VICS entities, IMPACT introduces realistic human-related uncertainties into the testing loop and supports the human-in-the-loop generation of corner cases as a complement to AI-driven methods. Furthermore, the mixedDT-enabled “Physical-Virtual Action Interaction” facilitates safe VICS testing under corner cases, incorporating real-world environments and entities rather than purely in simulation. Finally, we implement IMPACT on the I-VIT (Interactive Vehicle-Infrastructure Testbed), and experiments demonstrate its effectiveness. The experimental videos are available at our project website: https://dongjh20.github.io/IMPACT.
Accurate trajectory prediction of vulnerable road users is a cornerstone of safe autonomous driving and intelligent transportation systems. While large-scale pre-training has advanced this field, achieving robust zero-shot generalization remains a critical challenge for real-world deployment, particularly when vehicles encounter unseen environments and heterogeneous sensor configurations (e.g., varying frame rates and observation horizons). In this work, we revisit zero-shot trajectory prediction from the perspective of distribution shifts and distinguish three transfer settings: temporal transfer, scene transfer, and joint scene–temporal transfer. Through systematic experiments, we show that temporal mismatch is a key source of failure in current pre-trained models. By isolating temporal configuration from dataset shift, we demonstrate that explicitly conditioning on temporal metadata provides a simple and highly effective solution. Building on this insight, we propose OmniTraj, a Transformer-based framework pre-trained on large-scale heterogeneous data with explicit temporal-aware design. OmniTraj is designed to handle omni-generalization in trajectory prediction, namely adaptability across temporal configuration and scene shifts. It achieves state-of-the-art zero-shot generalization under joint scene–temporal transfer, reducing prediction error by over 70%. Furthermore, it exhibits exceptional robustness in safety-critical edge cases with severely limited observations and maintains high few-shot data efficiency, paving the way for scalable, dataset-agnostic deployment in real-world autonomous systems. The code is publicly available: https://github.com/vita-epfl/omnitraj.
This study presents a comprehensive analytical framework for rail system response performance, called the Generalized system Response Diagram model (GRD), which focuses on macroscopic dynamics and builds foundation of corridor-level performance characteristic analysis for rail systems. The GRD extracts traffic stream variables, including density, speed, and flow, from space-time train trajectories using Edie’s definitions, and identifies different rail traffic state regimes through critical points, defined by structural mathematical equations. This framework links corridor-level performance characteristics to underlying factors such as block signal control on interstation segments, interlocking constraints at stations and nominal maximum speed, providing a high-level perspective on rail system behavior under diverse operational states. Numerical experiments demonstrate the model’s ability to accurately represent rail traffic flow dynamics, supporting more adaptive and integrated rail traffic management strategies.
Origin-Destination (OD) flow, as a fundamental representation of urban population mobility, reflects the distribution of travel demand between spatial regions within a city. In intelligent transportation systems, precisely estimating city-level OD matrices is crucial for effective traffic management, formulating travel guidance strategies, and optimizing resource allocation. Existing OD estimation methods often rely on stringent traffic assumptions or exclusively leverage microscopic traffic observation to learn the mapping to OD flow. This dependency compromises their generalization capability in complex real-world urban scenarios. Consequently, a core challenge that remains to be addressed in this field is how to effectively model the traffic operation pattern and achieve high-accuracy OD estimation, especially under conditions where large-scale OD data collection is difficult. To address this, we propose a scalable framework with explicit mechanism-guided modeling (SEMM), which is designed to achieve high-accuracy estimation of urban OD matrices under conditions where large-scale OD data collection is difficult and labeled training samples are limited. The framework extracts multidimensional representations of operational patterns from the long-term operation of the transportation system from a macroscopic perspective, including time-domain difference, frequency-domain difference and stochastic perturbation term. Subsequently, the region-level relation equation explicitly models the mechanism-guided empirical relationship between regional production (i.e., departure volume) and average speed using these macroscopic representations. It effectively integrates microscopic traffic average speed information with macroscopic operation patterns, thereby enhancing the capability to capture traffic patterns and improving the expressive power of traffic status. Furthermore, we introduce a parameterized zero-inflated negative binomial (ZINB) distribution to precisely model the sparse distribution characteristic of OD matrices. Experimental results on two real-world datasets demonstrate that the proposed method significantly outperforms existing state-of-the-art baselines across various performance metrics as well as in inference overhead. Concurrently, the modular and scalable design enables SEMM to be seamlessly integrated with various optimization frameworks to leverage its dual advantages of high accuracy and efficiency. Our code is available at https://github.com/pku-smart-city/source_code/tree/main/SEMM.
Accurate and efficient human trajectory prediction is a fundamental capability for applications such as autonomous driving and intelligent transportation systems, requiring precise modeling of diverse social interactions and individual motion intention while meeting real-time inference demands. We present DID, a dual-condition flow matching framework that addresses these challenges through three complementary components. First, a density-adaptive dual-branch interaction encoder routes each scene between a lightweight spatio-temporal encoder for sparse settings and a high-capacity encoder combining physics-biased pairwise attention with learnable hypergraph reasoning for dense crowds, thereby avoiding spurious interactions in sparse scenes and insufficient group-level modeling in dense ones. Second, an asymmetric intention predictor distills scene context into a compact latent embedding via a prior branch, while a posterior branch, available only during training, provides oracle supervision. A posterior mixing mechanism mitigates the train-inference gap, enabling robust intention guidance without future trajectories at inference. Third, a dual-condition self-consistent flow matching decoder (DS-Flow) injects global intention via adaptive layer normalization and retrieves local interactions through cross-attention, while an EMA teacher enforces consistency between predictions along the interpolation path, enabling reliable single-step generation without iterative solvers. Extensive experiments show that DID achieves state-of-the-art or near state-of-the-art accuracy on the SDD and NBA datasets among generative methods while substantially improving inference speed, and delivers competitive results on ETH-UCY, confirming its effectiveness across diverse scene densities and its suitability for real-time deployment.
In this study, we introduce a novel cell-based dense parking service system to address the challenge of urban parking supply shortages. We propose two formulations, including a Round-based Formulation (RbF) and a Time-Space-Path-based Formulation (TS-PF), to produce the optimal parking schedule that minimizes the number of relocations caused by vehicle blocking. Both models can be easily extended to handle realistic variants. To address the limitations of solving these models individually, we propose a novel and general Alternative-Models-Induced Interactive Neighborhood Search Matheuristic (AMIINS). This approach effectively leverages the strengths of both RbF and TS-PF while circumventing their weaknesses, thereby significantly enhancing computational efficiency. By conducting extensive numerical experiments on real-world parking datasets, we show the substantial benefits of reservation consideration in reducing vehicle relocations and demonstrate that our AMIINS method significantly outperforms two benchmark methods in terms of solution quality, computational time, and convergence speed over instances of varying sizes (especially over practical-sized ones). Sensitivity analyses regarding initial solutions and relevant key algorithm parameters further demonstrate the robustness of AMIINS and offer useful tuning suggestions. Beyond the scope of this paper, the proposed AMIINS method is generalizable and can be extended to solve a wide range of optimization problems that have two or more alternative formulations.
Estimating traffic flow on unobserved roads is a critical yet challenging task for network-wide traffic state awareness. In urban settings, however, prevailing methods are often impractical due to their reliance on costly auxiliary data, or fundamentally inadequate because they model roads as spatial nodes-a practice that ignores the heterogeneous context shaping urban traffic dynamics. To bridge the gap, we propose a Knowledge Graph-based Pairwise Difference Learning Framework (KG-PDLF). First, we recast roads as semantic entities within a custom-designed Urban Traffic Context Knowledge Graph (UTCKG). It explicitly models the diverse factors and the distinct relationships of urban context that shape traffic dynamics, enabling the generation of context-aware road representations by knowledge graph embedding. Second, we propose a Pairwise Difference Learning (PDL)-based model tailored for unobserved-road estimation. Rather than targeting absolute traffic flows, this model exploits the statistical stability of cross-road flow differences, jointly estimating the corresponding pairwise objectives through a dynamic Multi-Task Learning (MTL) architecture conditioned on the observable network state. Empirical evaluation on two real-world urban datasets (Paris and Torino) showed that KG-PDLF achieved superior estimation accuracy and stable performance across scenarios with different proportions and deployment configurations of unobserved roads.
Clustering has been extensively studied in statistics and machine learning. This study introduces an operations research perspective by modeling and solving clustering within a two-stage stochastic programming framework. The formulation is motivated by their structural alignment: each object to be clustered can be treated as an empirical realization in the scenario set. We therefore model each observation as a second-stage scenario, while the cluster centers are shared first-stage decisions whose performance is evaluated across all scenarios. This perspective offers a clearer structural interpretation of clustering and allows domain knowledge to be explicitly encoded through the objective and constraints, thereby improving interpretability and empirical clustering performance. To assess the effectiveness of this approach, we apply it to usage behavior pattern mining (user segmentation) in subscription services. To handle realistic problem sizes, we introduce three model enhancements and a tailored column-generation matheuristic with an early termination rule (ET-CG), which shortens the long tail of column generation and significantly reduces runtime without compromising solution quality. We validate the proposed framework using six months of real-world parking subscription data from a commercial parking facility in Shenzhen, China. The proposed two-stage stochastic programming approach achieves average clustering loss reductions of 6.35% compared to K-medians, 7.95% compared to K-medoids, 5.74% compared to K-modes, and 6.42% compared to BanditPAM++. Additional experiments show that the model enhancements accelerate computation by a factor of 20.68 on average, and the ET-CG matheuristic achieves near-optimal solutions with high computational efficiency, solving real-world instances in 30.94 to 47.53 seconds (average 38.26 seconds). Beyond usage behavior pattern mining in subscription services, the proposed two-stage stochastic programming framework generalizes to other clustering problems that minimize an aggregated loss over the empirical distribution induced by the observed objects.
