ABSTRACT Rapid urbanisation and increasing vehicle density have intensified parking‐related challenges, leading to traffic congestion, travel delays, fuel consumption and environmental pollution. Although IoT‐enabled smart parking systems have been widely investigated, many existing approaches suffer from limited cross‐city generalisation, inadequate temporal modelling and poor interpretability. This study proposes an enhanced multilayer machine learning framework that integrates spatio‐temporal feature engineering, hybrid ensemble prediction and explainable artificial intelligence for parking occupancy prediction and parking behaviour analysis. The framework combines XGBoost and LightGBM to capture dynamic parking patterns while maintaining computational efficiency. A geographically separated multi‐city evaluation strategy is adopted using the SFpark (USA), Melbourne (Australia) and Birmingham (UK) datasets for training, validation and testing, respectively. A unified preprocessing and feature alignment framework is introduced to handle variations in spatial structure, temporal resolution and feature composition across heterogeneous datasets. Experimental results demonstrate that the proposed framework achieves an average accuracy of 96.1%, an F1‐score of 94.5%, RMSE of 0.121 and MAE of 0.095, outperforming conventional machine learning and deep learning models, including CNN‐LSTM. SHAP‐based explainability identifies hour‐of‐day, parking duration and day‐of‐week as the most influential features. The proposed framework provides an interpretable, accurate and generalisable solution for intelligent IoT‐enabled smart parking management.
ABSTRACT Traffic forecasting is essential for developing safer, more responsive and more efficient transportation systems. However, existing methods often exhibit limited temporal adaptability, inflexible spatial dependency modelling and insufficient horizon‐specific prediction refinement, particularly across heterogeneous highway and urban settings. To address these limitations, this study proposes MSMT, a memory‐augmented spatio‐temporal multi‐scale Transformer for network‐wide traffic forecasting. MSMT integrates four coordinated components: an adaptive temporal embedding module for contextual time representation, a multi‐scale temporal convolution module for capturing short‐, intermediate‐, and long‐range dependencies, a memory‐augmented spatial attention module for context‐conditioned historical retrieval and adaptive spatial reasoning, and a confidence‐guided multi‐resolution output head for horizon‐specific prediction refinement. Experiments on eight real‐world highway and urban datasets at 3‐, 6‐, and 12‐step horizons demonstrate that MSMT achieves the lowest MAE across all evaluated dataset–horizon combinations and provides the strongest overall performance across MAE, RMSE and MAPE. At the 12‐step horizon, MSMT improves average MAE and RMSE by 8.2% and 11.8%, respectively, relative to the strongest competing model. It also completes one training epoch in 58.60 s, maintaining practical efficiency. On PEMS04, MSMT requires 1.56 million trainable parameters, 0.99 GFLOPs per sample, and processes the complete test set in 12.60 s on an NVIDIA A40 GPU.
ABSTRACT Train Describer (TD) data provide a fine‐grained, event‐centric view of railway operations by recording signalling and infrastructure state changes. Although a subset of TD messages explicitly carries train identifiers, most signalling and track‐circuit (TC) events do not, making it difficult to determine which train triggered a given operational action. This limits the construction of train‐centric representations required for data‐driven railway analysis and decision support. This paper proposes a deterministic Cycle‐Constrained Train Association Framework (CC‐TAF) to reconstruct train‐aligned event sequences from raw TD data. The framework formulates train association as a structural reconstruction problem and adopts a hierarchical two‐stage strategy. At the signal level, heterogeneous TD events are organised into signalling cycles subject to lifecycle constraints, and CA messages serve as explicit anchors for inferring consistent train identities within each cycle. At the track level, train identities are further propagated to TC events through CA‐anchored temporal association under bounded search horizons. The proposed framework is designed to remain robust to practical data artefacts commonly observed in operational TD streams, including duplicated records, temporal disordering, incomplete cycles, and closely spaced train movements. Experiments on real‐world railway data from the Derby signalling area show that CC‐TAF produces a consistent train‐aligned TD representation and enables more effective integration of signalling data with train‐centric operational datasets. The resulting train‐level representation also improves downstream railway prediction tasks, including train delay prediction and route decision prediction. These findings demonstrate the value of deterministic train‐level reconstruction as a reliable preprocessing foundation for data‐driven railway analytics in safety‐critical environments.
ABSTRACT Long‐term traffic flow prediction remains challenging due to the time‐varying and non‐linear influence of meteorological conditions on traffic dynamics. This study proposes weather‐grouped temporal feature reconstruction (WGTFR), a forecast‐conditioned latent fusion framework that systematically incorporates weather forecast information into sequential and graph‐based prediction architectures. WGTFR comprises three coordinated modules: Dynamic feature weighting (DFW), which calibrates forecast‐aware weather relevance within latent representations; feature separation and directional fusion (FSDF), which disentangles strong and weak weather–traffic interactions and propagates them independently and hierarchical reconstruction fusion (HRF), which consolidates these interactions into prediction‐oriented latent states. Experiments on real‐world traffic and weather data from California evaluate WGTFR across four backbone architectures—Seq2Seq, Transformer, Autoformer and GCN‐Seq2Seq—under multiple long‐horizon prediction settings. WGTFR achieves the lowest RMSE across nearly all evaluated architectures and prediction horizons. Relative to weather‐free baselines, it reduces mean RMSE by 1.5% to 9.1% averaged across prediction lengths; relative to direct weather concatenation, a further 1.1% to 14% reduction is achieved. These results establish that forecast‐conditioned grouped reconstruction provides a reliable and architecture‐agnostic strategy for incorporating meteorological information into long‐horizon traffic forecasting, with performance gains increasing progressively with prediction horizon length.
ABSTRACT Mobile phone signalling data (MPSD) has become a valuable resource for driving numerous urban applications. However, due to high collection costs and strict privacy constraints, the scarcity of labelled datasets limits its broader application. To address this challenge, we reframe the problem from pure generation to a reality‐preserving simulation. This study proposes a deep reinforcement learning framework that synthesises MPSD trajectories by learning a policy to translate high‐fidelity GPS data into realistic cell tower connection sequences. The framework employs proximal policy optimisation (PPO) to learn a sophisticated connection policy, guided by a composite reward function that jointly optimises for point‐level accuracy and trajectory‐level distributional similarity, effectively capturing physical phenomena like connection hysteresis. A comprehensive evaluation across micro‐ (continuity) and macro‐ (distributional) levels demonstrates that our model outperforms state‐of‐the‐art baselines. The synthetic trajectories achieve high fidelity, with an effective point ratio of 0.97 (ground truth: 1.0), an average temporal sampling rate nearly identical to real data (79s vs. 77s), and low divergence in spatiotemporal distributions. Crucially, our approach enables lossless transfer of semantic labels (e.g., travel mode and purpose) from the source GPS data, directly addressing the label scarcity problem. The practical utility of the synthetic data is validated on two downstream tasks—map matching and travel mode identification—where models trained on our data achieve performance comparable to those trained on real data. This work presents a scalable solution to the data and label scarcity problem in mobility research, providing a robust method for generating semantically enriched, application‐ready MPSD for advancing large‐scale urban mobility analysis.
ABSTRACT The increasing integration of electric vehicles (EVs) provides new opportunities for fast frequency support, but also increases the cyber‐physical vulnerability of load frequency control (LFC). In communication‐assisted multi‐area systems, cyberattacks such as false data injection (FDI) and denial of service (DoS) can corrupt or block the signals used for secondary control, resulting in larger frequency excursions, degraded area control error (ACE) regulation and inefficient EV utilization. This paper proposes a cyber‐resilient LFC framework for EV‐integrated interconnected power systems, where EV aggregators provide both local vehicle‐to‐grid (V2G) support and coordinated cross‐area assistance under attacked conditions. An LSTM‐based detector is used to identify abnormal temporal behaviour and localize the attacked area. After detection, the compromised signal is replaced by a safe estimate, and healthy‐area EV fleets provide auxiliary support according to available power headroom and state‐of‐charge (SOC) constraints. The framework is validated using a two‐area system and a modified IEEE 39‐bus three‐area system under FDI, DoS, replay, stealthy FDI and multistage attack scenarios. The LSTM detector achieves a testing accuracy of 98.28%. In the modified IEEE 39‐bus system, the proposed mitigation reduces the frequency RMSE from 0.01508 to 0.01023 in Area 1 and from 0.02383 to 0.01539 in Area 2 compared with the unmitigated attack case, corresponding to reductions of 32.2% and 35.4%, respectively. The integral absolute error is also reduced by 35.2% in Area 1 and 50.3% in Area 2. The results demonstrate that combining attacked‐signal replacement with coordinated cross‐area V2G support improves frequency recovery, suppresses ACE deviations and maintains feasible EV operation under cyberattacks.
ABSTRACT Panoptic driving perception, which unifies object detection, lane line recognition and drivable area segmentation, is a cornerstone of autonomous driving. Multi‐task learning (MTL) is a natural paradigm to tackle these heterogeneous tasks jointly; however, conventional MTL frameworks often suffer from negative transfer, where conflicting task objectives lead to suboptimal feature representations and degraded accuracy. Such deficiencies pose safety risks in real‐world autonomous systems. In this paper, we propose YOLOP‐FCD, a novel multi‐task visual perception network that mitigates negative transfer through two key innovations. First, a progressive layered extraction module explicitly decouples task‐specific features while enabling selective feature sharing in a hierarchical manner, balancing common knowledge transfer with task isolation. Second, a synergistic multi‐attention transformer block integrates pixel‐wise, channel‐wise and spatial attention to enhance feature representation and cross‐task collaboration. Extensive experiments on the BDD100K dataset demonstrate that YOLOP‐FCD achieves competitive performance and a more balanced trade‐off among perception tasks. Ablation studies further validate the effectiveness of the proposed modules in alleviating feature interference and improving task synergy for autonomous driving perception.
ABSTRACT The coordinative use of ground vehicles (e.g., trains and trucks) and drones has emerged as a promising approach to enhance the efficiency and flexibility of emergency logistics. However, communication interference can severely undermine such coordination. To address this challenge, this study investigates a dynamic multimodal routing problem with drones under communication uncertainty in emergency logistics. The duration of the communication interference is considered an uncertain parameter. A mixed‐integer programming model is formulated with a bi‐level objective. The upper‐level objective is to ensure order fulfillment rate, and the lower‐level objective is to minimise cost. To address the computational challenges of this complex problem, a hybrid approach is developed, which integrates reinforcement learning (RL) and adaptive large neighbourhood search (ALNS) within an event‐triggered rolling‐horizon framework. In this approach, RL continuously learns from the real‐time data and guides the ALNS procedure to generate optimal routing decisions dynamically. The proposed approach is evaluated against a benchmark method, demonstrating superior effectiveness and adaptability. Furthermore, a series of experiments are conducted to analyse the effects of numbers of training iteration and the influence of different communication interference durations. The results confirm that the proposed approach improves the efficiency and reliability of emergency logistics under communication uncertainty.
ABSTRACT Existing traffic police gesture recognition methods often rely on computationally intensive 3D CNNs or RNNs using skeletal data. To address these efficiency challenges during feature extraction, this study proposes a two‐stage recognition framework based on optical flow enhancement. First, the RAFT algorithm extracts the motion vectors of gesture targets, suppressing background interference from vehicles and pedestrians. Second, a lightweight ShuffleNetV2‐K5 architecture is designed as the feature extractor. By employing 5 × 5 depthwise separable convolutions, the network expands its receptive field to capture large‐scale spatial features, thereby reducing the required network depth. For temporal modelling, the Sparrow Search Algorithm (SSA) is introduced to dynamically optimise the hyperparameters of an LSTM network, which helps reduce prediction errors and mitigate overfitting. Experimental results demonstrate that this optical flow‐enhanced lightweight model balances recognition accuracy with computational efficiency, shortening training time while maintaining stability. Ultimately, this framework explores the efficacy of optical flow features and provides a theoretical foundation alongside a potential algorithmic approach for real‐time gesture recognition under favourable lighting conditions.
ABSTRACT Recovering situation awareness (SA) from non‐driving related tasks during Level 3 takeovers is critical for safe and timely control transitions. However, out‐of‐the‐loop drivers often require an initial perceptual reorientation phase before effective SA recovery begins, a temporal delay that is frequently overlooked. Existing quantification models typically assume an instantaneous maximum recovery rate and therefore fail to capture the early temporal delay represented by the initialisation latency parameter, which may overestimate driver readiness during the critical early phase. This driving simulator study compared single‐ and two‐stage warning strategies and evaluated a parsimonious and interpretable sigmoid‐based dynamic modelling approach. Results indicated that two‐stage strategies produced an attentional priming effect, optimising visual attention distribution, enhancing subjective SA and improving takeover performance. Compared with the exponential reference model, the sigmoid model provided a better account of SA recovery dynamics. It characterised the delayed onset of SA accumulation in single‐stage recovery, whereas two‐stage strategies reduced the estimated initialisation latency. These findings suggest that graded warnings can serve as a cognitive scaffold that supports driver preparation before the control transition. This study provides a quantitative foundation for adaptive takeover warning design and supports the modelling of driver readiness dynamics in conditional automated driving.
ABSTRACT Rising energy prices and decarbonisation targets have increased interest in energy optimisation for urban rail systems, but earlier reviews focus on a single subsystem or method family and do not compare artificial‐intelligence‐based (AI) and non‐AI approaches in a consistent way. This paper presents a systematic literature review of 226 studies on energy optimisation in urban rail systems published between 2011 and 2025. Using the PRISMA protocol, we group the literature into five optimisation families: speed profile / energy‐efficient driving (SPF), timetable (TMB), regenerative energy braking (REB), energy management with storage (EMG) and traction power network (POW). Each family is distinguished between AI‐based and non‐AI approaches. Overall, 166 studies use non‐AI methods and 60 use AI‐based techniques. For single‐module problems, the median normalised energy‐saving indices are higher for non‐AI methods in SPF, TMB and EMG (18.7%, 18.51% and 15.22% versus 11.2%, 6.6% and 2.5% for AI). In contrast, for integrated categories that combine regenerative braking and storage, AI‐based methods achieve higher median savings, with REB + EMG showing 30.1% for AI versus 16.7% for non‐AI. However, integrated‐family medians are based on small samples and reported savings depend on heterogeneous baseline definitions. Therefore, the quantitative comparisons should be interpreted as preliminary, indicative evidence rather than definitive rankings.
Extracting topologically correct and kinematically feasible shipping routes from AIS data remains an open problem because existing methods formulate treating route extraction as a purely geometric fitting task and overlook the heavy-tailed, density-heterogeneous nature of maritime traffic. To address this gap, this study proposes a density-weighted graph search framework that re-formulates route extraction as least-resistance pathfinding on a density-weighted topology. The framework makes three contributions. (1) A navigable density field is constructed with a PowerNorm transformation to recover low-frequency oceanic corridors hidden by long-tail traffic distributions. (2) A tight-to-loose dynamic-radius topology suppresses land-crossing shortcuts while preserving global connectivity. (3) A heat-adsorption cost function combined with cubic B-spline smoothing and kinematic-aware semantic waypoint extraction yields decision-ready passage plans. Case studies on the Ningbo-Singapore and Singapore-Rotterdam corridors show that the framework achieves an RMSE of 4.8-5.2 km, a length ratio of 1.002 and a smoothness improvement of more than one order of magnitude over the baseline skeleton.
In this paper, a comprehensive method that integrates the energy optimization with the AC traction power supply modeling is proposed to achieve the energy-saving operation in the single-train and double-train scenarios. Concretely, a novel non-iterative traction power flow (TPF) method is introduced to reduce the computation time for time-varying load flow analysis. Meanwhile, a customized dynamic programming algorithm is designed to generate the optimal speed profile based on the TPF model. In the case of multiple high-speed trains, the optimization process enables the tracking train to maximize the utilization of regenerative braking energy (RBE) generated by the braking train. The energy cost function for the double-train coupled system integrates the substation supply and non-supply conditions to further enhance the energy efficiency. Finally, a real-world case study based on the Wuhan-Guangzhou high-speed railway is presented to demonstrate the effectiveness of proposed method. The results indicate that, compared with the traditional mechanical energy optimization schemes, the proposed method can reduce the total line-wide energy consumption by up to 2%, mitigating the catenary voltage fluctuations, and ensuring the precise power matching during the RBE transmission phases.
ABSTRACT Self‐driving vehicles (SDVs), also known as autonomous vehicles, are at the forefront of technological advancements in transportation, offering significant benefits in safety, efficiency, and mobility. However, these vehicles face critical security challenges due to their reliance on sensors, artificial intelligence (AI), and networked communications. This paper presents a comprehensive review of the security challenges in SDVs, including cyberattacks on vehicular networks, sensor spoofing, AI‐based threats, and privacy concerns. Unlike previous reviews, this study systematically categorises security challenges and solutions based on their nature and impact, and presents a detailed tabular taxonomy of current solutions to cybersecurity threats, a conceptual threat model for SDVs’ system layers, the state‐of‐the‐art of SDVs’ architecture, and their relative effectiveness. Furthermore, the latest advancements, such as AI‐driven intrusion detection, blockchain‐based security mechanisms, lightweight cryptographic protocols, and hybrid approaches, are explored. Finally, the review highlights existing research gaps and suggests future directions for enhancing SDV security.
The optimization of long-distance multi-objective operations for heavy-haul trains faces low solution efficiency when applying intelligent algorithms. Currently, few studies consider the changing emphasis of multiple operational objectives, such as safety, stability, punctuality and energy efficiency changes across different operational segments through task switching. To address this issue, this paper proposes a sequential multi-task collaboration hierarchical reinforcement learning framework for dynamic multi-objective optimization in long-distance operations. First, an event-driven time-dependent classification model is established to decompose the long-distance optimization problem into multiple subtasks, thereby reducing optimization complexity. Furthermore, a hierarchical reinforcement learning method based on sequential multi-task collaboration is designed, in which segment-specific subtask policies are trained in parallel and invoked sequentially along the route, while global coordination is achieved through a top-level dynamic programming framework. Simulation results based on a real 340-km railway line demonstrate that, compared to a single-agent global optimization method, the proposed framework improves convergence speed by nearly 40% and achieves superior optimization quality. It effectively restricts coupler forces within the safe limit of 1100 kN, while successfully balancing multiple objectives including speed tracking and energy consumption.
Detecting small objects in maritime scenes is critical for Maritime Autonomous Surface Ships (MASS) but remains difficult because vast low-entropy sea-sky backgrounds dilute the weak feature response of tiny targets in global vision transformers. To address this, the Recursive Signal Amplification Network (RSGN) is proposed. A Hierarchical Context-Scale Alignment (HCSA) stream couples a Swin Transformer backbone with multi-dimensional attention to align semantic features across scale, space, and channel. A Saliency-Guided Recursive Gating (SSRG) module then estimates a spatial saliency prior and recursively amplifies the signal-to-noise ratio of potential small targets before final detection. On the public WSODD benchmark, RSGN achieves 89.5% mAP and 36.3% APS, surpassing the baseline DINO by 6.9% and 5.9%, respectively.
Connected and automated vehicle (CAV) platoon control technologies have demonstrated significant potential in improving traffic efficiency, ensuring safety and reducing energy consumption. However, uncertainties in the driving environment can easily lead to control failures. Reinforcement learning (RL), with its strong learning and imitation capabilities, has been widely applied to CAV platoon control, yet it suffers from poor interpretability and output instability, posing safety risks. This study proposes a CAV platoon control method that integrates RL with physical information. A Twin Delayed Deep Deterministic Policy Gradient (TD3)-based platoon control model is developed. The design explicitly considers traffic efficiency, driving comfort and platoon stability. Physical rules are incorporated as an additional loss term into the policy network of this model, jointly guiding the policy iteration process of the actor network, thereby improving training efficiency and interpretability. The feasibility and effectiveness of the proposed method are validated under both low and high disturbance scenarios. The results indicate that, compared with physics-based and Model Predictive Control-based (MPC-based) methods, the proposed method achieves more balanced control under both disturbance conditions, with overall performance improvements of approximately 7.9% and 5.9%, respectively. The proposed method demonstrates excellent performance in vibration suppression, energy-efficient driving and generalisation capabilities, providing an effective technical solution for intelligent decision-making in complex driving environments.
ABSTRACT Lane changing is a critical behaviour affecting traffic safety and efficiency. Connected and automated vehicles enable multi‐vehicle cooperative optimisation in lane‐changing scenarios through vehicle‐to‐everything communication that facilitates inter‐vehicle state sharing and intention coordination. However, existing reviews lack a unified architecture classification framework for cooperative lane‐changing. Based on decision authority distribution and information interaction patterns, this paper categorises cooperative lane‐changing systems into two architectural types: centralised and distributed. It systematically reviews research progress across four core technical components: decision‐making, trajectory planning, motion control and system verification, establishing a comparative analysis framework for representative methods at each component. The paper reveals performance characteristics of different architectures and methods across multiple dimensions, including performance metrics, computational complexity and applicable scenarios, thereby providing researchers with a systematic conceptual framework. Furthermore, it identifies current technical challenges and outlines future research directions.
ABSTRACT Public transport users in many large cities must compare travel alternatives under uncertainty, especially when services lack schedules or fixed frequencies and some lines or modes provide real‐time arrival information within a limited reliability horizon. This problem becomes more complex when onboard crowding varies along the trip. This paper presents a routing algorithm for advanced traveler information systems to compute personalized shortest hyperpaths in transit networks with random arrivals, partial real‐time information, and real‐time or estimated onboard passenger counts. The network is modeled as a time‐dependent hypergraph with arc seat load factors varying between consecutive stops. The algorithm considers real‐time and random arrivals, transfer constraints, user‐determined seat load factor bounds, the user's haste value, penalty for riding, and crowding variability. It returns a Pareto‐optimal set of hyperpaths compared by expected travel time, transfers, penalty for riding, and crowding variability. Three options enforce the bounds by excluding arcs above the upper bound, allowing them only for a bounded time, or not enforcing them. An application example identifies hyperpaths of 30 and 35 min under partial real‐time information. Without enforcing the bounds, the fastest hyperpath takes 10 min, has no transfers, exceeds the upper bound, and has an infinite penalty for riding.
