Intelligent Transportation

Developing low-carbon transportation is crucial for ecological conservation and energy security. This paper addresses eco-driving challenges in connected vehicles by proposing a model-guided deep reinforcement learning framework to enhance the practicality and reliability of eco-driving strategies. Integrating model-guided and data-driven approaches, the method first designs a speed guidance algorithm based on the Intelligent Driver Model (IDM), which flexibly adjusts driving conservatism through parameter tuning. Subsequently, a reinforcement learning framework based on the Twin Delayed Deep Deterministic Policy Gradient (TD3) is proposed. The reference speed generated by the IDM is embedded into the state space as prior knowledge to guide the agent toward rapid convergence during training. This strategy significantly optimizes vehicle energy consumption while strictly ensuring driving safety and having minimal impact on travel efficiency. Experimental results demonstrate that compared to baseline methods and traditional deep reinforcement learning strategies, the proposed approach achieves superior performance in energy savings, driving smoothness, and traffic continuity. It achieves an average energy savings rate of 7.96% with only a 0.44% increase in travel time, validating its effectiveness and practicality for eco-driving tasks.

To enhance the energy and travel efficiency for electric bus routes, a cloud-supported efficient eco-driving control strategy based on a vehicle-cloud collaborative layered architecture is proposed in this paper. At the cloud layer, an efficient velocity planning model balancing energy consumption and travel time is constructed based on route map information and solved using the Dynamic Programming (DP) algorithm. At the vehicle layer, a safety-prioritized arbitration strategy is designed to switch decisions between the cloud-planned optimal velocity and the adaptive cruise control (ACC) car-following velocity to ensure driving safety. An experimental platform based on the vehicle-road-cloud collaborative architecture was constructed, and validation was conducted under real-world intercity road conditions. Experimental results demonstrate that the system exhibits favorable communication, real-time performance, and reliability. Compared with the ACC strategy, the proposed strategy increased average velocity by approximately 4.21% and reduced energy consumption by about 1.12% while ensuring operational punctuality.

Benefiting from the synergistic integration of aerial agility and terrestrial endurance, aerial-ground amphibious platforms can effectively traverse unstructured environments, demonstrating considerable potential for emergency response and reconnaissance applications. However, most existing systems adopt a decoupled configuration, utilizing independent propulsion units for ground and aerial modes. Although this approach enables dual-domain operation, the duplicated actuators and transmission chains inevitably incur structural redundancy and additional mass, which reduce overall system integration and payload efficiency. To address these limitations, this paper proposes a novel aerial-ground amphibious platform based on an integrated wheel-propeller structure. In the proposed design, a single electric motor actuates both the wheel and ducted propeller through a shared powertrain, where an electromagnetic clutch selectively engages the transmission path to switch between terrestrial locomotion and aerial propulsion. Building upon this mechanism, an amphibious platform configuration incorporating eight-wheel-propeller units is developed. By establishing a dynamic simulation model of the platform, the dynamic response characteristics during the drive-to-fly transition were investigated. The results demonstrate the feasibility of the operational principle underpinning the platform. The research provides a feasible technical approach for lightweight and highly integrated aerial-ground platforms, laying the foundation for future experimental implementation.

Large Language Models (LLMs) are propelling Intelligent Transportation Systems (ITS) toward an agentic paradigm, enabling intent understanding, constraint-aware reasoning, and decision-making. This paper focuses on the Group Meeting Location Recommendation (GMLR) problem, which aims to jointly optimize group preferences and mobility costs under real-world transportation network constraints. However, existing methods often overlook real-world geographic context, fail to jointly coordinate group preferences and spatial feasibility, and rely heavily on static, structured data to model group preferences, limiting their applicability to complex and dynamic group scenarios. We propose MEET-LLM, an agentic, LLM-driven, closed-loop framework for GMLR. MEET-LLM comprises three modules: (1) a Preference Profiling Module that constructs structured semantic representations of users and merchants by using LLMs to extract multilevel preferences from natural language reviews, overcoming the limitations of static aggregation; (2) a Spatial Optimization Module that introduces a mobility cost (MC) metric to evaluate group-level travel convenience and account for user-specific travel constraints; and (3) a Chain-of-Thought (CoT) Reasoning Module that coordinates group preferences and mobility costs through multi-turn LLM reasoning to generate interpretable, executable recommendations with map-based navigation guidance. MEET-LLM jointly coordinates group preferences and spatial feasibility, bridging natural language reasoning with real-world mobility execution. It enables dynamic, structured, and action-ready recommendations in real urban environments, highlighting the potential of LLMs as the cognitive core for agentic mobility coordination in ITS.

With the development of vehicle-to-infrastructure (V2I) and intelligent connected vehicle technologies, roadside sensing data provides a new foundation for vehicle queue prediction at signalized intersections. However, existing queue-prediction models still have prominent limitations. Mechanism-based models generally ignore the impact of signal-phase changes on driver start–stop behaviour and do not account for individual driver differences, while data-driven models act as black boxes and rely heavily on large-scale historical data. In addition, some connected-vehicle-based approaches are constrained by low penetration rates or limited real-time performance, restricting their engineering applicability. To address these issues, this paper proposes a microscopic driver-behaviour-oriented queue prediction method for signalized intersections based on roadside traffic data. The classical intelligent driver model (IDM) is extended by introducing a traffic-light remaining-time adjustment term, so that drivers’ anticipatory braking under red phases and speed adaptation near the end of green phases are explicitly embedded in the longitudinal acceleration. Together with the minimizing overall braking induced by lane changes (MOBIL) model, a unified prediction framework is constructed to describe both car-following and lane-changing decisions. Using vehicle position, speed, and signal-phase information collected by roadside sensors, a dynamic evolution model of queue formation, stagnation, and dissipation is further established, and quantitative estimation methods are developed for the maximum queue length and queue dissipation time. Multi-scenario validation is conducted on a simulation of urban mobility (SUMO)-based platform and with field data from the Yizhuang corridor in Beijing. The proposed method achieves a mean absolute percentage error (MAPE) of 20.66% in simulation and 25.73% in field verification, and outperforms conventional IDM-based and support vector regression (SVR)-based baselines in prediction accuracy across varying traffic demand levels. These findings indicate that the proposed method can provide practical support for signal timing optimisation, green-wave coordination, and V2I-based traffic management.

Multimodal Large Language Models (MLLMs) have achieved remarkable progress across a range of vision-language tasks and demonstrate strong potential for traffic accident understanding. However, existing MLLMs in this domain primarily focus on coarse-grained image-level or video-level comprehension and often struggle to handle fine-grained visual details or localized scene components, limiting their applicability in complex accident scenarios. To address these limitations, we propose SafePLUG, a novel framework that empowers MLLMs with both pixel-level understanding and temporal grounding for comprehensive traffic accident analysis. SafePLUG supports both arbitrary-shaped visual prompts for region-aware question answering and pixel-level segmentation based on language instructions, while also enabling the recognition of temporally anchored events in traffic accident scenarios. To advance the development of MLLMs for traffic accident understanding, we curate a new dataset, SafePLUG-Bench, which contains diverse multimodal question–answer pairs with detailed pixel-level annotations and temporal event boundaries across a wide range of accident scenarios. Experimental results show that SafePLUG achieves strong performance on multiple tasks, including region-based question answering, pixel-level segmentation, temporal event localization, and accident event understanding. These capabilities lay a foundation for a fine-grained understanding of complex traffic scenes, with the potential to improve driving safety and enhance situational awareness in smart transportation systems.

Timely detection of subsurface road defects is critical for structural safety and pavement longevity. Here, we propose MARD-Net, an enhanced deep learning framework designed for the accurate identification of urban subsurface road defects. First, addressing the scarcity of defect samples, a hybrid dataset was constructed by integrating empirical data acquired via the GS8000 ground-penetrating radar (GPR) system and synthetic data generated by gprMax. Second, to address complex geological backgrounds and variations in defect waveform scales, the RepNCSPELAN4_CAA module was integrated into the architecture. By combining structural re-parameterization with context anchor attention (CAA), this module enhances feature reuse and inter-channel interaction, enabling the fine-grained discrimination of subtle defects. Third, a lightweight asymmetric detection head (LADH), incorporating depthwise separable convolution (DSConv) within its regression branch, was developed to significantly reduce computational costs while maintaining robust detection performance. Finally, to overcome weak and uneven defect signals across imaging depths, a multi-path coordinate attention (MPCA) mechanism adaptively fuses global and local contextual information for precise defect recognition. Empirical experiments show that MARD-Net achieves a 2.07% increase in mean average precision at an intersection over union threshold of 0.50 (mAP@50) over the baseline YOLOv11n while reducing floating point operations (FLOPs) by 2.0 giga floating point operations (GFLOPs).

In traffic scenarios, the dynamic characteristics and random behaviors of vehicles are the main reasons for the frequent occurrence of collision accidents. Traditional early warning systems restrict traffic safety due to low detection accuracy and poor tracking effect. Research has proposed a vehicle safety distance early warning system based on deep learning to enhance traffic safety. Innovations include: adopting the self-calibrated illumination (SCI) algorithm to overcome light interference; the YOLOv11 algorithm is improved by introducing a secondary innovative cross-domain feature attention (CDFA) mechanism, reconstructing the feature pyramid, and integrating knowledge distillation to balance detection accuracy and real-time performance. The DeepSORT algorithm is improved by applying group convolution to reduce the number of parameters and replacing Intersection over Union (IoU) with MPDIoU to enhance tracking accuracy. The distance between vehicles is calculated by using the monocular vision ranging method. The detection, tracking, and ranging modules are integrated into a vehicle safety distance early warning system. Experimental evaluation demonstrates a marked improvement in the system's performance. On the public dataset, the detection model exhibits a gain of 3.89% in mAP@0.5 and 2.76% in mAP@0.5:0.95, while the tracking model achieves a 0.9% increase in multiple object tracking accuracy (MOTA). Furthermore, real-world vehicle validation confirms that the synergistic operation of the detection and tracking modules effectively mitigates the miss rate, thereby substantiating a tangible enhancement in overall system safety.

Flexible sensors have attracted wide attention like never before in the fast-growing flexible electronics era, because they are easily adaptable to curved or soft surfaces based on their inherent flexibility and simultaneously detect multiple external stimuli. In the field of intelligent driving, they utilize novel conductive materials, including conductive polymers, carbon-based nanomaterials, and liquid metals, to construct multifunctional flexible sensor networks, thereby accomplishing seamless integration with curved vehicle surfaces and comprehensive status monitoring. The advantages are that achieving dynamic deformation adaptability through flexible materials and manufacturing, enhancing system redundancy/robustness/real-time performance through a sensor network deployment, and improving perception accuracy through multimodal fusion. Therefore, flexible sensors exhibit great potential in intelligent cockpit interaction, chassis obstacle detection, and vehicle health diagnostics. However, its large-scale commercialization still faces challenges in automotive-grade integration, weather resistance, data security, and power supply. Furthermore, flexible sensors are expected to integrate with AI models, lightweight architectures, and self-healing smart materials in the future, thereby advancing autonomous driving development, facilitating vehicle-road-cloud coordination, and revolutionizing mobility paradigms.

For intelligent transportation systems (ITS), understanding pedestrian motion intention is crucial for enhancing traffic safety, enabling human-centered mobility services, and facilitating adaptive vehicle-pedestrian interactions. This paper proposes a pedestrian gait recognition method based on a modified particle swarm optimization-support vector machine (MPSO-SVM), utilizing fused surface electromyography (sEMG) signals and ankle joint angles. Seven lower-limb gait features were extracted from these signals to characterize walking patterns. The MPSO algorithm optimizes the support vector machine (SVM) parameters to improve classification performance. Experimental results based on data collected from healthy subjects demonstrate a recognition accuracy exceeding 92.5% across four gait phases. The proposed method offers significantly enhanced accuracy and robustness compared to traditional classifiers. These results suggest that the method is suitable for deployment in intelligent traffic control systems, autonomous vehicle navigation, and urban pedestrian behavior prediction.