ABSTRACT Passenger congestion in subway station waiting areas has emerged as a critical challenge in modern urban transit systems, particularly under high‐density conditions. Conventional object detection methods such as You Only Look Once (YOLO) often struggle in visually complex environments due to overlapping objects, occlusion and limited spatial resolution for precise object boundaries. To overcome these limitations, a semantic segmentation (SS)‐based framework is proposed for crowd congestion level classification, designed to estimate pixel‐level occupancy in closed‐circuit television (CCTV) imagery without relying on bounding boxes. Unlike conventional density estimation approaches that primarily focus on passenger counting, the proposed framework quantifies crowd conditions using occupancy ratios derived from pixel‐level segmentation results and classifies congestion levels based on space utilisation characteristics. Two additional correction models, U‐ASD Net and perspective‐aware convolutional neural network (PACNN), were incorporated to compensate for perspective distortion caused by varying distances and crowd densities within the scene, which result in significant variation in object scales across foreground and background regions. A dataset comprising 72,000 images per station from two major dense traffic transfer stations was analysed, and occupancy ratios were clustered into six levels of service (LOS) using a Gaussian Mixture Model (GMM). The SS‐based method demonstrated improved classification performance compared to the object detection baseline. Although correction‐integrated models showed slightly reduced average accuracy, their performance improved in severe congestion scenarios, indicating enhanced stability under visually degraded conditions. Statistical analysis confirmed significant model‐level variation, supporting the potential of segmentation and adaptive correction for practical application in transit surveillance environments. The proposed framework can support real‐time crowd monitoring and congestion management by providing quantitative occupancy‐based indicators for transit facility operations. Future work will focus on expanding the training dataset and improving model generalisation across diverse environments.
ABSTRACT During non‐peak periods, conventional buses (CB) often operate with low load factors, resulting in inefficient vehicle utilisation. To address this issue, this study develops a two‐stage sequential optimisation framework for coordinated operation of CB and demand‐responsive transit (DRT). In the first stage, DRT routing and scheduling are optimised using an improved genetic algorithm. In the second stage, a grid search algorithm determines the optimal CB departure interval for the remaining passenger demand. The framework is evaluated using Beijing bus IC card data and compared with the existing CB‐only operation mode. Results indicate that the proposed scheme improves operational efficiency, while the increase in passenger waiting time remains limited. In the upstream and downstream directions, total system costs decrease by 12.88% and 4.50%, respectively, while the average load factors of CB vehicles increase by 23.39% and 12.20%. The findings demonstrate the effectiveness of integrating DRT with conventional bus services and provide practical support for optimising public transport operations during non‐peak periods.
ABSTRACT Engine‐compartment thermal design requires economical prediction and physical assessment of candidate modifications. This study integrates field‐synergy‐guided input selection, support vector regression and Kalman‐smoothed grey‐wolf hyperparameter search for coolant outlet temperature and drive‐shaft sheath temperature . All 249 CFD cases are assessed using five outer and three inner folds, with scaling, covariance or EMA settings and SVR parameters selected within the inner procedure. KF‐GWO‐SVR achieves pooled out‐of‐fold RMSEs of 0.5678 and 1.6051, respectively. Equal‐budget comparisons with tuned EMA and nine‐start GWO show target‐dependent differences: KF has slightly higher errors and lower RMSEs than these two controls. The gain analysis characterises the transition from initial covariance‐dependent smoothing towards a steady EMA‐like weight. Frozen deployment models provide consistent conditional interaction surfaces, a finite‐candidate Pareto approximation and a selected design. Its rounded implementation is assessed against a matched geometry control in 12 independent paired vehicle‐test blocks, comprising 24 complete thermal cycles. Mean reductions are 1.183 for (95% CI 1.054–1.313) and 5.450 for (4.963–5.937), and every reported endpoint is below the project temperature limit. The frozen models predict larger reductions of 2.277 and 8.230. The study combines controlled predictive evaluation, explicit search‐mechanism analysis and a paired engineering assessment, with KF used as an optional search‐guidance strategy.
ABSTRACT Variable message signs (VMS), also referred to as dynamic or changeable message signs, are critical tools in contemporary traffic management, yet their effectiveness depends heavily on the design of displayed information and its interaction with driver cognition. While numerous studies have examined VMS impacts on driver behaviour, there remains a lack of structured synthesis linking design parameters, behavioural evidence and guideline development, particularly in developing‐country contexts. This scoping review addresses this gap by systematically analysing both international VMS design guidelines and empirical behavioural studies. Unlike prior reviews, this study (i) provides a parameter‐level comparison of eight national guidelines (e.g., font‐size, spacing, message phasing and units of information) and (ii) explicitly maps these design parameters to behavioural outcomes such as comprehension, gaze behaviour and reaction time. Using the Arksey and O'Malley framework, 1313 records were screened, with relevant studies synthesised alongside formal guidelines from diverse jurisdictions. Findings reveal strong convergence on key principles, such as message brevity and limited phasing, but significant inconsistencies in critical parameters, such as font size and information structuring. Empirical evidence consistently shows that excessive information load and suboptimal design increase cognitive demand, delay driver response and may compromise safety. Malaysia's current guideline is identified as lacking approximately 40% of design specifications present in comparable standards and does not adequately account for high‐speed driving conditions. The paper proposes an integrated evidence‐design‐policy framework to support the development of context‐sensitive, human‐centred VMS guidelines. The findings contribute both to advancing scientific understanding of the VMS design–behaviour relationship and to informing practical guideline improvements for Malaysia and other developing countries with similar intelligent transport system maturity.
ABSTRACT Accurate prediction of flight estimated time of arrival (ETA) and estimated elapsed time (EET) is essential for improving both safety and efficiency in air traffic flow management (ATFM). Existing flight time prediction approaches often underutilise prior knowledge embedded in flight plans, leading to physically implausible outputs and large prediction errors. To overcome these limitations, we propose FlightNet, a constrained deep learning framework that fuses flight‐plan semantics with kinematic priors for flight time prediction. FlightNet features a flight plan encoder with a hybrid embedding scheme to represent route structure and aircraft attributes, capturing spatiotemporal dependencies along the planned route. On top of the encoder, a dual‐branch decoder jointly predicts EET and ETA. To further enforce physical plausibility, we design kinematic constraint regularisation strategies (KCRS) that incorporate constraints such as time monotonicity and speed consistency into the training objective. Experiments on a real‐world dataset provided by the Civil Aviation Administration of China (CAAC) show that FlightNet consistently outperforms strong trajectory‐based and deep learning‐based baselines. Compared with OpenAP, FlightNet reduces mean absolute error (MAE) by 34.7% for EET and 36.4% for ETA; compared with TimeXer, MAE is reduced by 61.8% and 71.3% on the EET and ETA tasks, respectively. Ablation and analysis results further confirm the effectiveness of each proposed component.
