Ensuring the safety of electric vehicles (EVs) is a critical concern for next-generation mobility solutions, where the digital twin (DT) plays a vital role in addressing this challenge. This study provides a detailed analysis of the current state of knowledge in this domain. The findings highlight several effective applications of DT technology in enhancing EVs safety, particularly in battery maintenance and performance. However, our analysis reveals that most existing research is focused on battery-related applications, indicating the need to explore broader dimensions of EV safety. We identify key safety challenges, for example, high-voltage hazards, thermal runaway, system malfunction, and examine the role of DT in real-time monitoring and predictive maintenance to address them. We outline the DT-EVs enabling technologies, analyse the simulation platform, and present a unified DT-EVs architecture. This study classifies DT architectures and evaluates DT using artificial intelligence (AI) driven algorithms such as convolutional neural networks (CNNs), long short-term memory (LSTM), and generative adversarial networks (GANs), have achieved high accuracies in risk detection. By synthesising these insights, the objective of this study is to provide foundational conceptual support to researchers and, encourage the development of comprehensive, user-centric, and scalable DT-based safety solutions.
The logistics and transportation industry faces several critical challenges, including the absence of real-time vehicle data, inadequate safety measures, and inefficient resource management. This paper presents the Smart Logistic System (SLS), a unified intelligent transportation architecture that integrates embedded systems, Augmented Reality (AR), Machine Learning (ML), Internet of Things (IoT), and digital-twin technology within a single operational framework, rather than as independent subsystems. The system continuously monitors key vehicle parameters — location, speed, fuel consumption, and engine performance — to support proactive maintenance and minimize downtime. An edge-deployable driver-behavior-monitoring module employs computer vision and a lightweight MobileNetV3-Small classifier to detect drowsiness, distraction, and other hazardous behaviors in real time. Special-needs support includes seizure detection, GPS-based wandering alerts, and AR-guided vehicle-operation assistance. A route-optimization module formulates logistics planning as a constrained, multi-objective vehicle routing problem that jointly considers transportation cost, travel time, and estimated carbon emissions, while a digital-twin layer maintains a continuously synchronized virtual representation of fleet state to support fleet-manager decision-making. A mobile application delivers real-time data, alerts, and navigation to drivers; a desktop application allows managers to decode raw Controller Area Network (CAN) data; and a cloud platform provides secure storage, synchronization, and visualization. On a held-out benchmark, the driver-behavior-monitoring module achieves accuracy and inference-latency performance competitive with, and in several respects exceeding, recent state-of-the-art driver-monitoring methods; the route-optimization, predictive-maintenance, and digital-twin modules are validated through architectural and feasibility analysis rather than large-scale benchmarking. Taken together, these results indicate that SLS’s principal contribution is the integration of safety monitoring, logistics optimization, predictive maintenance, digital-twin synchronization, and accessibility features within a single operational architecture — a step toward more secure, efficient, and inclusive logistics and transportation systems.
Multi-agent cooperative trajectory planning in safety-critical systems faces safety assurance and scalable coordination challenges, especially under dynamic environmental disturbances. Traditional sampling and optimization-based methods lack real-time adaptability and do not scale well in multi-agent settings. Reinforcement learning methods show promise, but often suffer from safety constraint violations and sample inefficiency. This paper proposes MAF-DAC, a Multi-head Action-Filtered Deterministic Actor-Critic framework for safe multi-aircraft trajectory planning. MAF-DAC adopts centralized training with parameter sharing and decentralized execution. A shared actor-critic model is trained using experiences collected from all agents, and the trained actor is deployed to each aircraft agent for execution based only on local observations. A pre-trained Multi-head Action Filter Network (MAF-Net) serves as a differentiable safety filter to reduce unsafe and inefficient actions before execution. MAF-DAC is adapted to a multi-aircraft rerouting task under dynamically evolving thunderstorm scenarios with positional uncertainty. The simulation results show that the proposed method improves safety, scalability, and learning efficiency over representative baselines. It maintains 99% separation compliance, achieves a 95.5% task success rate under moderate uncertainty levels, and scales to 45 aircraft in a 30-minute window in the tested scenarios. These findings demonstrate the potential of MAF-DAC as a safe and scalable AI-driven decision support approach for tactical multi-aircraft rerouting in convective weather conditions.
Unmanned Aerial Vehicles (UAVs) are increasingly deployed as mobile nodes in intelligent transportation systems (ITS), facilitating traffic monitoring, emergency response, and urban air mobility. These applications demand precise modeling of UAV behavior and energy efficiency under diverse real-world traffic and environmental conditions. However, existing datasets often overlook crucial aspects such as UAV flight dynamics and multi-UAV collaboration in complex airspace. To bridge this gap, we introduce AMOVFLY, a comprehensive UAV flight status dataset captured from multiple AMOVLAB UAV platforms. AMOVFLY offers substantial improvements over existing datasets: agile heterogeneous devices, comprehensive and extended flight records, diverse flight conditions (including fixed and dynamic altitude/speed profiles), and synchronized multi-UAV flight data under uniform environmental conditions. These features provide a rich foundation for developing predictive models such as Temporal Convolutional Networks (TCN) to support energy consumption forecasting, multi-agent coordination, and adaptive flight control—key challenges in transportation-oriented UAV deployments, such as aerial logistics and traffic surveillance. Our evaluation further quantifies the fragility of static models against real-world dynamics. This finding underscores the critical need for online updates and robust trajectory planning to ensure reliable energy prediction and mission safety in dynamic ITS environments. AMOVFLY thus opens new avenues for improving UAV mission planning, energy prediction, and time-series forecasting in complex, real-world transportation environments. Dataset page can be found at https://github.com/YujiaoHu/AMOVFLY-Dataset
To guarantee a safe, efficient and resilient very low level airspace for diverse unmanned aerial vehicles, the U-space framework schedules trajectories in four dimensions- latitude, longitude, altitude and time - so that conflicts are avoided and capacity is maximized. In earlier work on the U-space service provider platform, conflicts were mitigated solely by time-shifting: requested flights were either authorized with a specific take-off slot inside the drone operator’s requested launch window or rejected when no conflict-free slot could be found. This paper proposes an enhanced mitigation scheme that widens the range of options available to drone operators whose flight plans would otherwise be rejected. Instead of a single shift, the new scheme searches for alternative requested launch windows whose positions minimize both delay and the probability of generating new conflicts. Three optimization models for locating these alternative windows are formulated and embedded in strategic deconfliction service. Their effectiveness is assessed through Monte-Carlo simulations that replay real flight track data over Castelldefels (Barcelona). Comparative metrics - including acceptance ratio, average delay and computational cost - demonstrate the trade-offs between the models and confirm that the proposed strategy significantly increases flight plans acceptance while preserving safety.
Environmental perception is fundamental to the safety and reliability of autonomous driving systems. Despite recent progress, existing frameworks often struggle to achieve multi-granularity fusion of BEV spatial features in complex traffic scenarios. Furthermore, some approaches treat perception tasks separately and do not adequately model inter-task associations. To address these challenges, we propose HDDPNet, a multi-task perception model for autonomous driving, which adopts an encoder-decoder architecture to enhance semantic modeling and cross-task collaboration. In the encoding stage, we introduce a Hierarchical Cascade Encoder (HCE), which incorporates a TransMamba module to enhance global context modeling and feature representation. In addition, a Progressive Feature Pyramid (PFP) module is designed to facilitate multi-scale and fine-grained information fusion. In the decoding stage, we develop an Isomorphic Dual-stream Diffusion Decoder Network (ID3Net), which adopts a diffusion-inspired single-step denoising process to improve feature robustness and promote cross-task collaboration by sharing semantic and geometric information within a dual-stream architecture. Experimental results show that the proposed method achieves significant improvements in multiple perception tasks: NDS is increased by 10.09% in 3D object detection; AMOTA is improved by 7.29% in multi-object tracking; and mAP is increased by 2.17% in online mapping. These results demonstrate that HDDPNet consistently improves multi-task perception performance, providing a reliable solution for autonomous driving.
Environment perception based on multimodal data fusion is crucial for the development of autonomous driving (AD) technology. However, the existing fusion-based detection and tracking methods often struggle to efficiently utilize the multimodal complementary information. This limitation can significantly impact the safety of autonomous vehicles (AVs), especially in complex traffic scenarios and adverse weather conditions. To address this issue, this work proposes a universal multimodal data fusion framework called HybridFusion for 3-D object detection and tracking tasks. Specifically, a novel Modal-Specific Feature Interaction Strategy (MSFIS) is introduced within the detection algorithm. The strategy complementarily fuses multimodal features while preserving the potential information from individual modalities. Furthermore, a Motion-Appearance Coupling Driven Data Association Strategy (MACD-DAS) is incorporated into the tracking algorithm. The strategy comprehensively processes spatial motion information and multimodal appearance features, effectively reducing the uncertainty in state prediction and feature matching. Extensive experiments were conducted on View-of-Delft (VoD) and K-Radar datasets. The experimental results demonstrate that the proposed HybridFusion significantly improves the performance of multimodal detection and tracking algorithms in complex traffic scenes and adverse weather conditions. The demo is available at https://youtu.be/el_5jtxDO2I
Multimodal transit networks with strong structural interdependence are highly susceptible to cascading failures that spread rapidly during disruptions. Although existing studies have improved the understanding of failure propagation, the approach to mitigate the impact of cascading failures from the perspective of passenger redistribution remains insufficiently examined. In our study, a bi-level modelling framework is constructed with its upper level focusing on modeling the spatio-temporal dynamics of cascading failures. The lower level develops a Heterogeneity-Aware Physics-Informed Deep Learning Network (H-PIDL) to allocate redistributed passenger flow. The framework makes full use of resilient nodes and redistribute passenger flow to mitigate the impact of cascading failures. A multi-scenario case study is conducted on three representative metro-bus double layer networks in Nanjing, China, including a transportation hub area, a dense urban center, and a sparse suburban area, using real-world passenger flow and network data. Comparative analyses with traditional redistribution strategies demonstrate that the proposed framework moderates early escalation, reduces spatial spread, and enhances network performance. These findings highlight the value of redistribution considering node heterogeneity for mitigating the impact of cascading failures in multimodal transit systems.
Road damage is a major threat to transportation safety, and an effective pavement hazard detection system is important for road maintenance and transportation. Recent methods for detecting pavement distress may have issues with large network structures, including high computational complexity, insufficient feature extraction, and inaccurate bounding-box regression, making them unsuitable for deployment on embedded systems. In this study, a lightweight real-time pavement distress monitoring system based on cloud–edge collaboration is proposed. The system utilizes LEEP-YOLO11 based on YOLO11n, which is deployed on edge devices to enable real-time distress detection and store uploaded distress information through edge computing modules and cloud systems. First, in the YOLO11n network backbone, four LCNet (Lightweight Context Network) blocks are used, two for multiscale feature extraction and two for connecting to the neck layer. Then, we introduced EnMCA (Enhanced Multidimensional Collaborative Attention) that incorporates a Dynamic Context Awareness (DCA) module to reduce parameters and improve feature representations, dynamically adjust contextual importance, and combine with the original MCA module for improving the adaptability of the model to complex scenes. Furthermore, an EnMPDIoU (Enhanced Multipoint Distance IoU) is introduced to improve bounding-box regression accuracy, while a pruning strategy removes redundant filters to boost computational efficiency. Finally, a knowledge distillation procedure is conducted to further improve the detection accuracy. Experimental results demonstrate that LEEP-YOLO11 achieves a 6.15% increase in mAP@0.5 while reducing GFLOPs by 20.4% and parameter count by 23.1% compared to YOLO11n. These improvements make it suitable for deployment on embedded devices.
This study investigates the string stability of a novel, coastal passenger-freight traffic flow mixed with human- and intelligence-driven vehicles on single-lane container port freeways. The mixed flow considered herein comprises scheduled connected and automated truck platoons (CATPs), human-driven cars (HDCs) and container trucks (HDTs). The intelligent driver model (IDM), Adaptive Cruise Control (ACC) and Cooperative Adaptive Cruise Control (CACC) models are utilized to model the car-following behaviors of human-driven cars and trucks, and truck leaders and followers in CATPs, respectively. Stability discriminant equations and numerical simulations are adopted as analytic methodologies. Experimental results show thatthehomogeneous flow of ACC trucks remains stable with the expected time headway larger than a threshold (2.45 s), and the homogeneous flow of CACC trucks with shorter control cycles and larger headways tends to be more stable. All else being equal, as the CATP platoon lengths or the HDC penetration rates increase, the stability of the heterogeneous traffic flow improves gradually. Managerial implications are proposed to improve the stability of the mixed traffic flow considered in this study on container port freeways.
The rational matching of vehicle, control, and track parameters is critical for enhancing the ride comfort and stability of a 600 km/h electromagnetic suspension (EMS) high-speed Maglev system and facilitating its engineering application. To address this issue, this study investigates the sensitivity of key parameters influencing the system dynamic responses and conducts multi-objective parameter matching optimization based on the quantified sensitivity results. First, the stability conditions of the minimum suspension unit are derived, and resonance avoidance analysis is performed to exclude unstable parameter combinations. Second, a data-driven surrogate model integrating deep neural network (DNN) and Gaussian process (GP) methods is trained to enable rapid large-scale computations. Then, global sensitivity analysis is conducted using the Sobol method. Finally, the quantified sensitivity results are used to guide a multi-objective optimization process, yielding an optimized parameter set. The computational results provide quantitative sensitivity coefficients for the effects of individual parameters on the system responses. Under the optimized parameters, compared with the initial parameters, the standard deviation of the suspension gap is reduced by 15.45%, and the ride comfort index is reduced by 10.93%. The findings of this study provide theoretical support and strategic guidance for parameter matching in 600 km/h high-speed Maglev systems.
This paper investigates a min-max routing problem with time windows (RTW), which seeks paths for multiple agents starting from a single depot to visit a set of target locations, each within a designated time window, while minimizing the maximum completion time among the agents. Most existing research either considers the min-sum objective (minimizing the sum of completion times) or ignores the time window constraints. Recent work explores data-driven approaches, such as reinforcement learning (RL), to address this challenge. However, it still struggles to generalize to unseen scenarios outside the training set. To address these challenges, we propose a neuro-symbolic method called NS-RTW, which combines learning-based and classical search techniques within a bi-level optimization framework, enabling end-to-end self-supervision and offering better generalization than existing methods. On synthetic datasets, our method is able to find feasible solutions for all instances, outperforming the RL baseline, which fails in complex scenarios. Among all baseline methods, ours finds the best-quality solution in 87.1% of commonly solved instances. Experiments on Solomon benchmark and real-world case further demonstrate the generalization capability and practical applicability of our method.
Automotive radar is a key sensor for autonomous driving perception because it provides all-weather robustness at low cost. Among radar data representations, radio-frequency (RF) data has attracted increasing attention because it preserves rich target and scene information. However, single-frame RF-based detection methods cannot exploit temporal context and therefore often fail when target spatial features are weak. Although multi-frame methods introduce temporal information, most of them learn spatial and temporal features within the same extraction process, which leads to mutual interference and limits the quality of spatio-temporal feature representations. To address this issue, we propose a spatio-temporal dual-branch fusion network (SDFNet), which employs two dedicated branches to extract spatial and temporal features separately, alleviating mutual interference caused by jointly learning both types of features in the same process. Subsequently, a spatio-temporal feature fusion module complementarily fuses these two extracted features to generate comprehensive representations containing both fine-grained spatial details and dynamic temporal cues, thereby improving detection robustness and accuracy. In addition, we replace the standard max pooling operation in the temporal branch with a direction-aware pooling module, which reduces directional information loss during downsampling and better preserves direction-dependent temporal information. Experimental results demonstrate that SDFNet outperforms representative single-frame and multi-frame methods.
Adaptive transit signal priority (TSP) and bus control systems have been independently developed to maintain efficient and reliable urban bus services. However, those two systems could potentially lead to conflicting decisions due to the lack of coordination. Although some studies explore integrated control strategies based on deterministic optimization, route-level coordination of multiple traffic signals and buses cannot be maintained considering the underlying randomness in bus operations, which could result in suboptimal decisions and severe performance deterioration due to model mismatch in real-world traffic environments. In this study, we propose a robust route-level integrated control strategy based on hierarchical stochastic optimization to enhance bus schedule adherence under stochastic environment. Within the hierarchical framework, the route-level coordination and stochasticity modeling is decoupled, where the upper model jointly schedules multi-cycle signal plan and multi-bus arrivals along the entire route to provide global guidances, while the lower models handles local uncertainties with stochastic programming to enhance robustness, thus efficiently handles stochastic disturbances with joint optimization of route-level schedule adherence. Simulation results under various scenarios reveal that our approach significantly enhances bus schedule adherence with a mild impacts of 1.1%-3.3% on car delays, demonstrating the effectiveness of route-level coordination and high robustness under stochasticity. In addition, a field experiment conducted on sandbox testbed further validate the applicability of our approach in real-world control systems.
The Internet of Vehicles (IoV), as a key component of smart cities, plays a pivotal role in vehicle coordination and intelligent transportation systems. However, with the rapid advancement of IoV technology, the growing threat of network attacks has heightened the demand for enhanced security and privacy protection. Owing to the heterogeneous nature of devices within the IoV, which often employ different cryptographic systems, traditional authentication mechanisms face limitations in verifying the legitimacy of such heterogeneous devices. To address this challenge, this paper proposes a privacy protection and traceability anonymous mutual authentication protocol for heterogeneous vehicular networks (PPT-AMA). PPT-AMA can realize mutual identity verification and session key agreement between the vehicle in the certificateless cryptosystem and the roadside unit in the identity-based cryptosystem to ensure the secure transmission of private data. Moreover, by providing traceable pseudonyms and anonymous authentication for each vehicle, PPT-AMA protects vehicle privacy while preventing attackers from tracking vehicle information through public information such as public keys or identities. Under the discrete logarithm and computational Diffie-Hellman assumption, PPT-AMA is shown to meet various security requirements, including perfect forwards secrecy and known session-specific temporary information security. The performance evaluation, which is based on theoretical analysis and simulation experiments, demonstrates approximately 32.50% improvement in computational efficiency over existing schemes, highlighting the protocol’s potential for practical application.
In human-ADS co-driving scenarios, driver trust is critical for achieving efficient and safe human–automation collaboration which underlines the importance of accurately modeling its dynamic evolution. However, existing trust models lack refined consideration of autonomous driving scenarios and integrated modeling of both psychological and behavioral trust. To address the above issues, this study develops a human-ADS co-driving trust model from a cognitive process perspective. Grounded in Bayesian decision theory, the model is guided by interrelations among task-decision confidence, psychological trust and behavioral trust, and accounts for influencing factors such as driving scenario complexity and ADS’s ability. To fit and validate the model, a human-in-the-loop simulated driving experiment is conducted. The results show that the proposed model can characterize the dynamic evolution of psychological trust in human-ADS co-driving scenarios. In addition, this study introduces a personalized psychological trust threshold to explain the mapping from psychological trust to behavioral trust. The threshold-based prediction achieves an average behavioral trust prediction accuracy of 89.54% on the validation dataset, showing good robustness and providing a reference for trust calibration.
High-precision and fast tracking of displacement and velocity targets is critical to the safe and efficient operation of heavy-haul trains. However, in complex operating environments, traction motors acting as actuators are prone to saturation and failures (e.g.,torque saturation and oscillatory output), which may induce longitudinal impulses between adjacent cars and pose a severe threat to running safety. To address these issues, this paper proposes an anti-saturation adaptive sliding mode robust control algorithm that ensures tracking precision even under actuator saturation and failures by preemptively compensating for saturation deviations and tolerating unforeseen faults. First, a dynamic model of heavy-haul trains incorporating actuator saturation and failures is established, and an adaptive fast sliding mode observer is designed to estimate both actuator fault values and composite resistances in real time. Second, a pre-activated anti-saturation compensator is developed, which engages to eliminate saturation deviations within finite time when the control input approaches the actuator’s physical limits. Third, an anti-saturation fast sliding mode robust control strategy is proposed, featuring a nonsingular terminal sliding manifold and an integral feedback structure to avoid control singularities and mitigate chattering. Lyapunov stability analysis establishes the finite-time convergence of the global state errors, ensuring accurate tracking of the desired target by the heavy-haul train. Ultimately, MATLAB/RT-LAB co-simulations validate the proposed algorithm’s advantages in convergence speed, robustness, and control precision.
This paper addresses the coordination problem of Connected and Automated Vehicles (CAVs) at unsignalized intersections in mixed traffic with Human-Driven Vehicles (HDVs). The proposed formulation targets traffic-efficient and energy-efficient objectives, while ensuring safety, thus resulting into a Mixed-Integer Nonlinear Program (MINLP), which is hardly solved in reasonable time. Hence a computationally-efficient heuristic is proposed in this paper, which relies on the predictor-corrector method from sensitivity analysis tools. We perform numerical simulations to evaluate the economic and computational performance of the proposed Sensitivity-based Heuristic (SH) against MINLP, First-Come, First-Serve (FCFS), and an Exact Heuristic (EH) and show that SH can reach close-to-economic optimal solutions while consistently being computationally fast across different mixed-traffic scenarios.
Extreme weather disrupts transportation networks through abrupt capacity loss, evacuation surges, emergency-vehicle conflicts, damaged sensing, and conflicting field reports. Large language models (LLMs) can help operators interpret heterogeneous text and policy constraints, but using a large model at every control interval is slow, carbon-intensive, and unsafe without independent verification. This paper proposes carbon-aware verified model cascades (CA-VMC) for climate-resilient emergency traffic control. CA-VMC combines a spatio-temporal graph predictor, a risk-triggered router, small/large language-model planning branches, and a deterministic verifier that projects candidate plans into admissible signal-route-transit actions before actuation. The formulation distinguishes vehicle and electric-grid transportation emissions from AI-serving emissions, so model calls are optimized as part of a cyber-physical-carbon budget. The revised evaluation is specified as a Houston flood-response protocol: a SUMO-ready OpenStreetMap network is coupled with public flood and precipitation products, shelters, hospitals, vulnerable zones, demand generation, signal timing, fleet emissions, hardware profile, prompt schema, latency, and carbon factors. Across 96 paired simulation episodes, CA-VMC reduces mean evacuation time from 126.4 to 79.8 min relative to fixed-time control and from 91.0 to 79.8 min relative to DCRNN+MPC, while reducing transport CO2 from 8.42 to 6.08 tCO2 per episode. Relative to always-on LLM control, it reduces AI inference CO2e from 2.71 to 0.39 kgCO2e while preserving zero deployed safety violations. The results indicate that LLM-enabled traffic AI should be evaluated as a bounded, verified control stack rather than as a free-form planning interface.
Car-following behavior is the result of combined effects of various types of factors in real transportation systems. However, there is no car-following model that comprehensively incorporates the combined impacts of micro and macro factors, as well as the differences in car-following behavior among diverse drivers under the same impacts, i.e., driver heterogeneity. This paper presents a panoramic situation expression-based car-following model. First, the micro factors (including the types and positions of the vehicles in the current lane and adjacent lanes) and the macro factor (including the density of traffic flow ahead) that affect car-following behavior are modeled as the panoramic situation, and based on nonlinear mechanics and phase-field coupling theory, the mathematized expression of panoramic situation is established to achieve a unified expression of the combined impacts. Second, a car-following behavior decision-making method is proposed to incorporate driver heterogeneity based on cumulative prospect theory and the “decision-execution” mechanism of car-following behavior. Third, intelligent driving model is employed as the execution framework and integrated with the decision-making method to establish a car-following model. Real and virtual vehicle driving experiments are conducted to collect data and complete calibration. The fitting capability, numerical along with behavioral soundness, and scenario robustness of the proposed model are tested, evaluated, and verified through comprehensive utilization of the collected data, NGSIM dataset, and data of supplemental driving experiments in 15 cases within 10 representative scenarios. The capability of the proposed model to capture heterogeneous drivers’ car-following behavior, the cognitive rationality of its outputs for heterogeneous drivers, and the contribution of each module in the model to performance improvement were also tested.
