This study aims to model driver cognitive states for shared-control vehicles operating within intelligent transportation systems. We introduce driving activity level as an interpretable state variable that combines attention allocation, steering behavior, and perceived situational pressure for adaptive authority allocation. A driving simulator experiment with 12 licensed drivers was organized as a four-by-four factorial design involving distraction tasks, pressure tasks, and crossover conditions. Eye-movement measures, steering-derived variables, and a perceived-pressure model were integrated through principal component analysis to construct activity and performance indicators, and a feedforward neural network was trained to estimate activity level from gaze and pressure-related features. The results show an inverted U-shaped relationship between driving activity and performance: drivers performed best at moderate activity levels, whereas insufficient engagement and excessive pressure were associated with poorer control outcomes. The trained model produced consistent activity estimates in manual-driving and shared-control driving tests. These findings indicate that driving activity level can serve as a practical driver-state variable for real-time monitoring and adaptive authority allocation in intelligent transportation systems.
Efficient highway merging is essential for traffic safety and efficiency. The rise of connected and automated vehicles (CAVs) offers new opportunities to enhance merging operations. This study introduces a novel control algorithm that optimizes CAV trajectories within a hierarchical hybrid control framework consisting of three layers: strategic, tactical, and operational. At the top layer, the system identifies vehicles most affected by merging and relays this data to the tactical layer. There, a multiphase kinematic model and a multi-input multi-output (MIMO) model predictive control (MPC) scheme generate optimal trajectories. The operational layer uses these trajectories to guide vehicle maneuvers: longitudinal movement is controlled by a Proportional-Integral (PI) controller, while lateral movement is managed using a Proportional-Integral-Derivative (PID) controller, both are tailored to the specific dynamics of each vehicle. Simulations across various scenarios demonstrate the algorithm’s reliability, robustness, and accuracy under diverse initial conditions. The paper also proposes a centralized control technique that integrates all three layers into a unified, continuous control process. When compared, the hybrid control approach offers superior computational efficiency and resilience to model uncertainties and communication disturbances, while the centralized control achieves better stability, fuel efficiency, and passenger comfort.
Accurate and rapid traffic flow prediction is crucial for traffic management and the planning of intelligent transportation systems. With the increasing complexity of traffic networks and the widespread deployment of sensors, the scale of traffic flow data has significantly increased. Prediction models designed for small-scale datasets often struggle when applied to large-scale traffic networks. They typically encounter challenges such as long training times and high computational resource consumption. This limits the scalability of existing models. To address these issues, this study proposes a traffic flow prediction model based on a Multi-Subgraph Spatio-Temporal Graph Convolutional Network (MuSTGCN). Our model divides the original graph into multiple subgraphs by identifying pivotal nodes. We employ spatio-temporal graph convolution modules for each subgraph to capture spatio-temporal features. The results of these subgraphs are integrated through a parallel output module to generate the final prediction result. Furthermore, we developed a phased training strategy to train each subgraph separately, thereby improving efficiency. Experiments conducted on real-world traffic flow datasets demonstrate that the MuSTGCN achieves high prediction accuracy on large-scale datasets containing 8600 nodes. Additionally, our model significantly reduces training time, inference time, and GPU memory usage. This study offers a novel approach to developing prediction models for large-scale traffic flow data.
This study presents a machine learning framework for classifying recurring and non-recurring traffic congestion using threshold-based alert logs from Florida's Traffic Sensor Subsystem (TSS), addressing the research gap for methods that use only alert logs and remain computationally feasible for large-scale real-time deployment. One year of statewide TSS alerts (July 2022-June 2023) on Florida's limited-access highways were collected, with this dataset containing 692,420 alerts from 1,873 sensors. Alerts were labeled as non-recurring congestion by associating them with Florida Highway Patrol computer-aided dispatch incident records using a 60-minute temporal window and 1-mile spatial buffer. Using only alert timestamps and coordinates, a set of engineered features was constructed to characterize alert frequency, timing irregularity, and behavioral deviations. These features were used to train an XGBoost classifier that differentiates between recurring and non-recurring congestion without requiring external data sources such as crash reports or weather feeds. The model achieved an accuracy of 0.646 and a precision of 0.229, improving the existing baseline TSS precision of 0.127 by 80.3% and reducing false alerts reported to traffic management center operators by 39.4% for the statewide test dataset. Comparable performance was observed when the model was applied to a separate test dataset for Florida's Turnpike Enterprise roadways only. Temporal analysis showed that the model's true positive rate was highest between 9:00 AM and 10:00 PM on weekdays, offering practical guidance for traffic management center operations. This approach has great potential to be utilized by traffic operators in a real-time setting.
Modeling driver behavior in car-following scenarios is crucial for traffic simulation and the development of intelligent transportation systems. In dense urban environments, particularly in stop-and-go traffic, capturing the nuances of individual driving styles is essential for realistic predictions. This paper addresses the limitations of the well-established Optimal Velocity Model (OVM) that fails to identify the drivers’ behavior and yields physically unrealistic parameters. A new driver model is developed via collecting detailed vehicle motion data from stop-and-go urban traffic using image processing. We introduce a speed policy-based driver model via the fitting of a universal differential equation to the data. Then, the identified policy is approximated by a simple, piecewise smooth function with a small number of physically interpretable parameters. These parameters and the driver reaction times are identified from the traffic data with excellent consistency. Using the principal component analysis, driving behaviors are characterized. Finally, numerical simulations of vehicle strings confirm that the proposed model is more applicable and robust for simulating stop-and-go traffic than the traditional OVM.
Motion planning for autonomous vehicles is a pivotal component in realizing autonomous driving. However, at unsignalized intersections, vehicular mutual perceptual occlusion creates occluded areas where the lack of information about potential vehicles (PVs) within these areas can compromise the safety of autonomous vehicles. To address this, we propose the State Estimation-based Occlusion-Aware Motion Planning (SEOAMP) method, comprising two modules: the PVs multimodal motion state estimation module and the Efficient Occlusion-Aware Partially Observable Markov Decision Process (EOA-POMDP) motion planning module. The PVs motion estimation module first uses a probabilistic occupancy inference (POI) model to infer road occupancy ahead of the occluding vehicle, utilizing its spatiotemporal information. Subsequently, it fuses the inferred occupancy information using evidential theory to indirectly estimate occupancy within the occluded area, thereby estimating the multimodal motion states of PVs. The EOA-POMDP explicitly incorporates multimodal probabilistic state estimation of PVs into the partially observable state space, integrates various occupancy possibilities in the occluded area, and balances smoothness and efficiency when planning safe vehicle motion. Additionally, we employ a human-inspired virtual-lane-enhanced policy tree that reduces the computational complexity arising from PVs multimodal motion state while introducing extra virtual lanes for risk avoidance. Simulation results show that the proposed motion planning method enhances both driving safety and traffic efficiency for autonomous vehicles in unsignalized intersection scenarios with occlusions.
Personalized lane-changing trajectory planning is crucial for enhancing the safety and comfort of autonomous vehicles. However, existing methods typically incorporate personalization only at the final planning stage, failing to consider intrinsic differences in behavior mechanisms across driving styles. Consequently, the planning trajectory may still deviate from actual driver expectations, potentially causing human-machine conflicts that compromise comfort and safety. Therefore, this paper first quantifies behavior heterogeneity and analyzes demand differences across driving styles by naturalistic driving trajectories. Based on this, a personalized lane-changing trajectory planning framework (PLTPF) that includes personalized feasibility domain construction (PFDC) and personalized lane-changing trajectory optimization (PLTO) is proposed. In the PFDC, stylized lane-changing demand thresholds, vehicle dynamics constraints, and vehicle interaction collision boundaries are embedded to construct a personalized feasibility domain. It can accurately identify the personalized initial planning solution space and feasible trajectories tailored to different styles. In the PLTO, a personalized lane-changing cost function is established, incorporating driving preferences including driving risk, efficiency, and comfort. An adaptive differential evolution algorithm with dangerous parameter shielding (DPS-ADE) is deployed to select an optimal style-adaptive trajectory. Extracting real vehicle trajectories from the HighD dataset to construct lane-changing scenarios, experimental results show that the proposed planning framework can generate safe, comfortable, and personalized lane-changing trajectories for drivers with different styles.
Certain types of driving behaviors (e.g., searching or cruising for a parking space) may contribute to a series of urban issues, including excessive searching for parking and traffic congestion, among others. While some studies have analyzed the impact of various factors on cruising times, fewer studies have developed methods for detecting different parking-related behaviors from trajectory datasets. In this research, using detailed driving behaviors captured by passively collected in-vehicle GPS trajectory data, we introduce a computing framework that extracts parking trajectories from driving trajectories and classifies them into cruising or direct parking categories. The framework consists of a preprocessing module, a rule-based distributed parking trajectory extraction module, and an LSTM-autoencoder deep clustering model. The preprocessing and parking trajectory extraction modules are designed to extract trajectories that involve parking from raw driving trajectories. The deep clustering model is designed to take feature-augmented parking trajectories of varying lengths, without labels, as input and return two broad categories representing cruising and direct parking. The model returned two types of cruising and two types of direct parking. The results reveal spatial differences in parking behaviors, with more cruising concentrated in commercial and entertainment areas of Washington, DC. This study contributes a scalable framework for extracting and classifying unlabeled parking trajectories from big GPS trajectory data. The framework can be applied to trajectory data collected for any city and will be useful for identifying locations in an urban environment with cruising times that could be reduced to ease traffic flow and parking demand.
Convolutional Neural Networks (CNNs) have played a pivotal role in driver behavior recognition but often neglect global feature dependencies when stacking layers. Although transformers alleviate this issue, their high training costs hinder practical deployment. To address this challenge, we propose the Retention Vision Transformer (RVT), a lightweight architecture designed to reduce training overhead. Built upon the stage-wise design of TWINS, RVT incorporates a retention mechanism with Rotational Position Encoding (RPE), enabling parallel attention computation through a hidden state matrix. We evaluate RVT on the publicly available State Farm dataset, comparing it with several representative vision transformers and the lightweight CNN MobileNetV3. RVT achieves over 99% accuracy with only 10.1 M parameters and 0.7GB memory usage, reducing training memory consumption by approximately 80%, while exhibiting faster convergence. Although each training step incurs slightly higher computational cost than MobileNetV3, RVT converges within about 1/10 of the training epochs, resulting in lower overall computational requirements. These results demonstrate that RVT effectively reduces image size requirements, parameter count, computational overhead, and memory usage, while maintaining state-of-the-art accuracy. The proposed method thus holds strong potential for real-time driver behavior recognition and related applications.
This article introduces a novel compensation approach for actuator delays in a homogeneous Cooperative Adaptive Cruise Control (CACC) vehicle platoon. The Constant Time Headway (CTH) spacing policy is adopted, and a Predecessor-Follower (PF) communication topology is assumed. Vehicle dynamics are modeled with a third-order system that accounts for large, unknown actuator delays. To address these delays, we propose a compensation method based on the delay-independent Truncated Predictor Feedback (TPF). Specifically, a low-gain TPF-based controller is designed to achieve stabilization without requiring explicit knowledge of the actuator delay. It is rigorously proven that the proposed approach ensures both individual and string stability, maintaining zero steady-state error, regardless of the delay magnitude. Simulations and a comparative study with a baseline CTH-CACC controller are conducted to assess the theoretical findings. Results show that the proposed TPF-based controller compensates for significantly larger delay with neither its exact value, nor a delay estimate is required in the controller design.
Disruption operation is a critical safety factor in urban rail transit (URT). To address the challenge of train operation adjustments during unidirectional operation disruptions, we propose an adjustment strategy of "reverse operation & short-turning". With a focus on minimizing passenger waiting time at stations, we develop a two-stage MILP model for train operation adjustment under unidirectional operation disruption in the disruption and the recovery phases. An improved adaptive neighborhood search algorithm, tailored to the model's characteristics, is designed for optimization. The case study shows that under the unidirectional operation disruption, the passengers' waiting time organized by the "Reverse Operation and Short-turning Integration" (ROSI) strategy is reduced by 5.19% compared with that by the reverse operation strategy. Additionally, this strategy decreases the waiting time of passengers on the whole line by + approximately 19.80%, compared to closing the disrupted direction line, resulting in more balanced transportation services in both directions. Furthermore, when the disruption exceeds 15 mins, the proposed strategy outperforms the approach of closing the disrupted direction line in terms of operational efficiency.
Recent advances in sustainable mobility have accelerated autonomous driving technologies, particularly for public transport convoys of mechanically connected vehicles, improving the mobility and safety of urban mobility. This work develops a control architecture for such convoys, focusing on trajectory tracking and vehicle coordination. Conducted within a UniFi research project on automated electric minibuses, the study evaluates both mechanical coupling and conventional platooning strategies to optimize the reliability of urban transport. A Nonlinear Model Predictive Controller (NMPC) is designed for vehicle lateral and longitudinal control, ensuring precise tracking while considering actuator constraints and vehicle dynamics. The approach is first applied to a single vehicle and then extended to a two-vehicle convoy, where a Proportional-Integral (PI) controller manages the distance between vehicles. The follower replicates the leader's path while avoiding corner cutting and maintaining a predefined gap. The simulation results demonstrate excellent performance under ideal conditions, achieving precise trajectory tracking with a distance error limited to +/- 0.05m . However, Model-In-the-Loop (MIL) tests reveal a larger deviation (up to 0.80 m ) due to trajectory characteristics and dynamic constraints. The MATLAB/Simulink (R) implementation, validated with GNSS-acquired (Global Navigation Satellite System), confirms the viability of the proposed strategy, paving the way for further optimisations and experimental validations.
Situational awareness and decision-making are critical to aviation safety. This work investigates the feasibility of an intelligent task assistance approach based on natural language processing to assess situational awareness and support decision-making. At runtime, prompts are dynamically constructed by semantically pairing transcribed pilot utterances with thematically grouped operational variables—such as airspeed, altitude, and weather conditions—using sentence embeddings and structured contextual grounding. The proposed method evaluates whether spoken statements and decisions are supported by explicitly defined operational context, enabling the detection of semantic inconsistencies and the generation of data-grounded feedback. The approach was evaluated under controlled experimental conditions using manually constructed utterances representing typical flight scenarios and decision statements, including variations with transcription-like noise. Results indicate stable and deterministic inference behavior within this constrained setting, as well as sensitivity to decoding configuration, input quality, and contextual alignment. Model outputs remained grounded in the structured operational information explicitly provided in the prompts. These findings demonstrate the feasibility of semantic grounding and contextual inference using large language models as an architectural component for situational awareness assessment and support their potential use as assistive tools for aviation applications.
In autonomous driving scenarios, accurate semantic segmentation is crucial for environment perception and decision-making. However, traditional segmentation methods typically assume that the distribution of training and testing data remains consistent. In real-world driving environments, variations in weather, lighting conditions, and geographic differences often cause significant domain shifts, leading to substantial performance degradation of the models. Moreover, the existence of long-tailed class distributions means that certain critical objects appear infrequently in the training set, making it challenging for models to effectively learn the features of these rare categories, thereby affecting overall segmentation performance. To address these challenges, this paper proposes a novel domain-generalized semantic segmentation method that integrates an adaptive class sampling (ACS) strategy with a super pixel-guided hierarchical transformer decoder (SPHTD), enhancing the model’s robustness and generalization under distribution shifts. First, we introduce an ACS mechanism that dynamically adjusts the sampling frequency of training samples based on class statistics, ensuring sufficient representation of rare categories during training. Second, a low-rank fine-tuned visual foundation model is adopted as the network backbone to enhance feature extraction capabilities. Third, we design the SPHTD, which clusters pixels to generate localized masks and then integrates them into a coherent global representation through cross-attention, improving upon conventional pixel-wise classification. Finally, we conduct extensive experiments across multiple challenging cross-domain settings. The results demonstrate that our method achieves superior domain generalization performance, outperforming existing methods in both synthetic-to-real and real-to-real scenarios, while maintaining robust segmentation accuracy under adverse weather conditions.
Traffic flow at signalized intersections can be extremely unbalanced, and the phase time for the vehicle movements with low flow could be quite short. That may cause green time waste against minimal green time or significant time lost from phase switch. To address this problem, this article introduces the concept of sub-cycles, which are incorporated into pre-timed signal control to skip some vehicle movements in some sub-cycles, which are released in the others. Specifically, a lane-based pre-timed signal timing method with sub-cycles is proposed as a mixed-integer linear optimization model, where road capacity is maximized. To improve computational efficiency, all the feasible lane marking schemes are first enumerated, and then evenly assigned to the computer cores for parallel computing where optimal signal timing is proposed. Case study shows that the proposed model manages to enhance road capacity in most of the cases (4% on average and 7% at most) in all the unsaturated cases. Furthermore, scheme analyses demonstrate that the proposed model for lane-based pre-timed signal timing with sub-cycles is capable of enhancing the saturation of the movements originally with extremely low saturation. The research provides insights on applying pre-timed signal control to the intersections with unbalanced traffic flow, to promote signal timing efficiency.
In intelligent transportation systems, accurate cumulative arrival curves are essential for calculating key performance metrics of signalized intersections, such as vehicle delay and queue length. Recording vehicle identities and timestamps departed from approaching lanes at intersections, License Plate Recognition (LPR) systems are desirable to provide refined time-dependent arrival curves. Existing studies have estimated arrival curves by assuming cyclic arrival rates with the exact knowledge of signal timing data. This assumption restricts the model’s flexibility in characterizing arrival differences across signal cycles, especially under varying signal timing plans. In addition, the complete signal timing data across the network is not always accessible in practical applications. This study proposes a Periodic and Constrained Gaussian Process (PCGP) model that leverages LPR data to estimate lane-based cumulative arrival curves. Instead of assuming specific arrival patterns, a variable periodic kernel is designed to automatically capture the varying periodicity of arrival curves under varying signal timing plans without depending on signal timing data. Additionally, a monotonicity constraint is imposed to ensure the monotonically increasing characteristic of the cumulative arrival curve. In experiments, we analyze the model performance under varying signal timing plans and different function lanes (e.g., dedicated and shared lanes) and inspect the significance of model components. The results demonstrate the effectiveness of the proposed model in capturing varying periodicity in arrival processes at signalized intersections.
Vehicular ad-hoc networks (VANETs) are highly dynamic systems composed of vehicles and infrastructure nodes such as Roadside Units (RSUs). Secure participation and data sharing in VANETs are essential for safe integration in smart-city infrastructure. Incentive-driven data exchange can significantly strengthen Mobility as a Service (MaaS) platforms, including ride-sharing, smart transit systems, and real-time navigation services. This article presents an end-to-end VANET incentive framework that integrates adaptive trust management with a node-to-node handshake protocol for secure message exchange and validation. The proposed system rewards participating nodes for exchanging VANET messages, such as Basic Safety Messages (BSMs), and provides solutions to mitigate malicious behaviors, particularly collusion between nodes and unreliable data reporting. Nodes aggregate digitally signed handshake records into datasets, referred to as queues, which are submitted for validation and reward allocation through a lightweight private blockchain architecture. The framework incorporates trajectory tracking and dynamic adjustment of acceptable queue sizes based on regional traffic conditions and network behavior. This adaptive thresholding mechanism improves trust management by reducing unnecessary blockchain transactions and prioritizing data quality. The system was evaluated using exploratory simulations and replicated statistical batch experiments with real-world mobility datasets. Simulation results indicated that adaptive thresholding substantially reduces upload activity and communication overhead compared with fixed-threshold approaches while maintaining reliable participation behavior. The proposed framework demonstrates the potential of adaptive trust management and lightweight blockchain architectures to support scalable data sharing in connected transportation systems and other decentralized environments involving highly mobile nodes.
Road markings are an essential component of the modern transportation system. During the road operation period, clear markings are important for the autonomous lane-keeping system and dynamic traffic management. Modern transportation planning and maintenance practices highlight the significance of an accurate program to investigate marking conditions. Recently, computer vision has become prevalent in monitoring transportation infrastructures. However, the real-world variability brings substantial challenges to deep learning programs. Based on Unmanned Aerial Vehicle (UAV) images, we propose Vision Foundation Model (VFM)-based approach for robust marking damage detection and evaluation. First, an image selection method is designed to find representative instances. Self-distillation with No Labels Version Two (DINOv2), Modified Hausdorff Distance, Uniform Manifold Approximation and Projection (UMAP) and K-means clustering are performed to understand the underlying trend within UAV images. Second, we conduct You Only Look Once (YOLO) to localize damages. Third, the semantic segmentation is launched to handle road markings. Without pixel-level annotations, we carry out Segment Anything Model (SAM) to estimate the initial state of markings. Fourth, our program performs a thresholding step to calculate the missing ratio and severity of damaged markings. We test the proposed method in practical applications. The computation results indicate that our VFM-based method exhibits competitive capabilities in training efficiency, prediction accuracy and generalization ability. Based on the image selection program, YOLO significantly improves detection accuracy and mitigates training cost. Our SAM-driven program removes the annotation step and handles image variations. Considering its robustness and generalization, VFM-based method has potential for various future applications.
This article introduces an approach for signal predictive optimization utilizing real-time detection-based vehicle trajectory data. Unlike traditional trajectory data collected from probe vehicles, the detection-based vehicle trajectories are acquired by integrated radar and video detectors strategically deployed at signalized intersections. The detection comprehensively covers entry and exit lanes around the intersection with nearly complete penetration, thus being able to collect and process trajectory data in real time. A lane-level traffic state representation model is proposed to map vehicle trajectories collected in real time to individual lanes, facilitating the description of traffic flow dynamics. The optimization model is formulated as a mixed-integer linear programming problem with the objective of throughput maximization. Simulation experiments validate the effectiveness of the proposed strategies for network control, demonstrating advantages over fixed-time coordinated schemes, actuated control, and pressure-based adaptive control. While decision-making at each intersection is independent, a certain coordination effect is observed. Sensitivity analyses reveal the performance of the proposed strategy under various demand scenarios, objective functions, and detection ranges. This study explores the application of advanced real-time detection-based trajectory data for urban traffic management.
Short-term origin-destination (OD) prediction for urban rail transit is critical for efficient operation and dynamic management. Unlike conventional passenger flow forecasts that focus solely on aggregate indicators such as entries and exits, OD prediction reconstructs the complete station-to-station flow matrix, enabling fine-grained operational planning and providing a stronger basis for service design, transfer coordination, and capacity allocation from a network-wide demand perspective. However, the high dimensionality and sparsity of the OD matrix, along with complex spatiotemporal dependencies, make this task challenging. This paper introduces multi-dimensional holiday and weather tags, which are coupled with historical OD tensors into a unified multi-dimensional tensor input. The dual-graph spatial encoder captures the physical network topology and latent functional relationships between stations, while the hierarchical time-series Transformer models the short-term dynamics and periodic patterns. Furthermore, we integrate Tucker tensor decomposition in the decoder to efficiently reconstruct the OD prediction results and address data sparsity. Experiments on large-scale urban rail transit datasets show that Spatio-Temporal Graph Transformer Network (STGTN) significantly outperforms benchmark methods in prediction accuracy, with a 58% reduction in Mean Absolute Error (MAE) and an 11% increase in accuracy. Larger relative errors are mainly concentrated in peripheral or infrequently used OD pairs, whereas major travel corridors are predicted accurately. Ablation studies further confirm the contribution of each module and the external factors. These results demonstrate the effectiveness of tensor-based spatiotemporal modeling and highlight the practical value of OD prediction for short-term capacity deployment, transfer coordination, and congestion mitigation in urban rail transit systems.
