IET Intelligent Transport Systems最新文献

To improve the management and operational efficiency of Intelligent Transportation Systems (ITS), address the nonlinear complexity of short-term traffic flow, mitigate the issue of significant noise in traffic flow datasets, and tackle the challenges in determining parameters for Support Vector Machines (SVM) and Long Short-Term Memory (LSTM) networks, this paper proposes a short-term traffic flow prediction model based on Variational Mode Decomposition (VMD) and Least Squares Support Vector Machine (LSSVM) integrated with an attention mechanism. Multiple intrinsic mode functions (IMFs) decomposed by VMD are input into the LSSVM model, and the parameters and weights of the model are automatically adjusted using the attention mechanism. Experimental results on the Italian highway traffic flow dataset show that the prediction accuracy of the VMD-LSSVM-Attention model is improved by an average of about 38.6% compared with the traditional VMD-SVM, VMD-LSTM-Attention, VMD-LSSVM and LSSVM-Attention models, and the model is more stable. Furthermore, in generalisation validation experiments on the Rotterdam and Madrid datasets, the model improved prediction accuracy by 5.17% to 20.97% compared to the best-performing advanced models. This model provides a prediction method for the traffic flow prediction module in the intelligent transportation system (ITS) architecture.

This paper presents an integrated hybrid reinforcement learning–model predictive control (RLMPC) framework for autonomous highway systems, unifying macroscopic traffic flow regulation and microscopic on-ramp merging control. At the macroscopic level, a ramp metering (RM) controller based on a data-driven model predictive control (MPC) formulation using second-order Q-learning is implemented in the METANET environment on a benchmark three-segment freeway without the need for explicit traffic models. The RLMPC RM learns optimal flow regulation directly from closed-loop data, achieving enhanced system performance, constraint satisfaction and smooth control compared to common RM algorithms such as ALINEA, MPC and deep RL. At the microscopic level, an RLMPC merging controller manages autonomous on-ramp manoeuvres in which an ego vehicle enters the mainline approximately 160 m before the merge point and completes the manoeuvre 50 m downstream while interacting with surrounding vehicles. In this phase, when a collision risk arises, the MPC takes control; otherwise, the reinforcement learning (RL) policy operates, combining model-based safety with learning-based efficiency and yielding superior overall performance. Evaluations under varied traffic conditions show that implementing RM at the macroscopic level significantly improves microscopic on-ramp merging performance. Relative to the no-RM baseline, the framework achieves a 34.5% reduction in merge time under slow traffic conditions, eliminates collision events and moderately enhances overall efficiency and driving comfort.

To address the issues of missed detections, false positives and feature degradation in road debris detection caused by small and irregular targets, this study proposes a novel framework integrating multi-scale feature fusion and dynamic feature enhancement mechanisms. It also constructs a dedicated road debris dataset to fill the gap in public benchmark datasets in this field. Firstly, a cross-layer connection-optimized feature fusion network is designed in the neck network, addressing the limitation of insufficient fusion of shallow and deep features in existing methods, realizing efficient linkage between shallow texture features and deep semantic information, and significantly improving the detection capability for small targets. Secondly, a context-aware anchor attention module integrating reparameterized convolution and adaptive weight allocation is embedded into the backbone network. Compared with traditional fixed receptive field convolution, it can dynamically enhance target features and suppress background interference, effectively solving the problem of feature degradation in complex environments. Thirdly, an improved spatial pyramid fast pooling module based on global pooling and Ghost convolution is proposed, overcoming the defect of prone detail loss in traditional max-pooling and preserving key information of small-sized road debris to the greatest extent. Finally, a weighted fusion loss function integrating corner distance loss, focal loss, cross-scale correlation loss and CIoU loss is designed, breaking the limitation of insufficient attention to irregular targets in a single loss function and enhancing the model's adaptability to complex scenes. Experimental results show that the framework outperforms existing mainstream methods in road debris detection scenarios, achieving a precision of 91.5%, a recall of 82.0% and an mAP50 of 88.7%.

Cooperative adaptive cruise control, also known as vehicular platooning, has gained significant interest for its ability to enhance fuel efficiency and comfort in vehicle operations. This study proposes novel control strategies for vehicular platooning based on long short-term memory (LSTM) neural networks. By learning temporal dependencies in vehicle behaviour, the proposed LSTM-based controllers improve string stability within the platoon, particularly under varying velocity patterns of the lead vehicle. Two distinct frameworks are investigated: centralized and decentralized control models. The centralized model makes use of the states of all vehicles within the platoon, whereas the decentralized model focuses on the states of only a limited number of preceding vehicles. Simulation experiments demonstrate that both the centralized and decentralized LSTM controllers significantly outperform traditional, non-LSTM-based controllers in minimizing cumulative inter-vehicle error. This study contributes a novel controller training methodology that integrates LSTM-based architectures with optimal control principles, offering improved adaptability and flexibility in real-time platoon management.

Road intersections are the key nodes in urban road networks, and their operational efficiency affects the dynamic balance of regional traffic flow. However, classical intersection delay models are unable to quickly and efficiently predict large-scale intersection delays. This paper proposes a spatiotemporal graph convolutional network (STGCN) intersection delay prediction model based on delay big data. By introducing a temporal attention mechanism, the model can capture key temporal information and enhance the ability to model the spatio-temporal evolution law of intersection delays. First, based on traffic platform delay big data, external factor data, and road network structure, the historical delay matrix, the external factor matrix, and intersection adjacency matrix are constructed, respectively. Next, the graph convolutional network (GCN) is used to extract spatial features of intersection delay. A temporal attention mechanism is then introduced to assign weights to different time steps, thereby enhancing the model's perception of critical time steps. Finally, the temporal convolutional network (TCN) is employed to extract temporal features of intersection delay, which are used to predict future intersection delays. Experimental results show that, compared with the optimal benchmark model STGCN, the proposed model reduces the MAE, RMSE, and MAPE metrics by 3.80%, 2.75%, and 4.23%, respectively. This study on intersection delay prediction using a STGCN based on delay big data not only improves the efficiency of intersection delay prediction but also provides a theoretical basis for traffic management departments to formulate measures.This study focuses on the delay prediction problem of intersections, key nodes in urban road networks. Aiming at the deficiency that classical models are difficult to efficiently predict the delays of large-scale intersections, a STGCN model based on delay big data is proposed. This model constructs the historical delay matrix, the external factor matrix and the intersection adjacency matrix, uses the GCN to extract spatial features, combines the temporal attention mechanism to enhance the perception of key periods, and then uses the TCN to mine temporal features to achieve delay prediction. Experiments show that, compared with the optimal benchmark model STGCN, its MAE, RMSE, and MAPE indicators are reduced by 3.80%, 2.75%, and 4.23%, respectively, providing theoretical support for improving the efficiency of intersection delay prediction and traffic management decision-making.

Off-road path planning and navigation often struggle with complex challenges, such as diverse surface conditions that demand adaptability, stability-sensitive vehicle dynamics on low-adhesion terrain, and the persistent trade-off between real-time performance and path quality. To address these challenges, an improved rapidly-exploring random tree (IRRT) algorithm is developed to adjust the dynamic exploration domain considering the vehicle's design speed and local terrain features, which can affect vehicle's operational stability, thereby increasing path feasibility and environmental adaptability. Furthermore, a nonlinear model predictive controller (NMPC) is deployed in the lower layer of the proposed RRT path planning framework, smoothing the generated path and enhancing ride comfort through terrain-aware adjustments. Both a 100 × 100 meter simulated environment and a real-world 1:10 scale test site, featuring distinct terrain types, i.e., hard roads, natural terrain, and low hills, with obstacles. The results show that the proposed two-layer path planning framework, improved RRT algorithm integrating with NMPC, reduces path length by 6.9% and total turning angle by 12.3% compared to RRT, while maintaining a maximum curvature of 0.134 m⁻¹ (well within the safety limit of 0.2 m⁻¹) and improving ride comfort by 80.4%. On the other hand, although the computation time increases by 272.2%, the resulting gains in path quality and stability justify the trade-off. The proposed method demonstrates a viable solution for off-road vehicle navigation across diverse terrains, effectively balancing path feasibility, ride smoothness, and computational efficiency.

Current approaches to intercity dynamic ride-sharing mainly adopt single-stage scheduling, where new orders are periodically batched and processed. Although effective, this strategy often causes heavy computation and delayed passenger feedback, limiting real-time applicability. To address these issues, we propose a novel two-stage information feedback framework combining coarse and fine scheduling. In the coarse stage, online scheduling (nearest insertion) promptly matches new orders with departed vehicles, while offline scheduling (best insertion) processes non-departed vehicles, thus providing passengers with staged and timely feedback. In the fine stage, assignments are further optimised through large neighbourhood search, with the triggering decision modelled as a Markov decision process and learned by deep Q-learning. This design reduces redundant computation while dynamically balancing feedback timeliness and scheduling efficiency. Unlike traditional methods, our framework is novel in integrating staged passenger feedback, hybrid heuristic optimisation and reinforcement learning-based control. Experiments on two real-world intercity carpooling datasets show that the method significantly reduces runtime and feedback delays while maintaining strong scheduling performance, demonstrating its potential as a practical solution for large-scale dynamic ride-sharing platforms.

Urban traffic prediction is of great significance for traffic management and optimisation. Although research on predicting indicators such as traffic flow and speed is relatively sufficient, research on forecasting congestion ratios in different urban regions remains inadequate. Based on traffic big data, this paper proposes a fusion regional congestion ratio prediction model integrating eXtreme gradient boosting tree (XGBoost) and long short-term memory (LSTM), which integrates multi-source features, including temporal, meteorological, and spatial factors. First, the XGBoost algorithm is used to model the historical congestion ratios and related features of each region, obtaining preliminary prediction results and extracting regional residual sequences; subsequently, the residual sequences are input into the LSTM network for error correction. Finally, the prediction results of the two stages are fused to obtain more refined regional congestion ratio predictions. Experimental results show that during peak hours on weekdays, taking Region 49 as an example, the MAE of the fusion model is 0.062, the mean absolute percentage error is below 30%, and the comprehensive prediction accuracy reaches up to 72%; under complex weather conditions, for the same region, the RMSE values of the fusion model are 0.048, 0.058, and 0.043, respectively, which are 37%–63% lower than those of the XGBoost model used alone. Feature ablation experiments further verify the key role of temporal, meteorological, and spatial features in improving prediction performance, among which spatial features contribute the most to performance optimisation. This study improves the research framework in the field of urban traffic prediction and provides a theoretical basis and methodological support for regional traffic management practices.

This paper proposes a novel Signal-Vehicle Cooperative Control framework (SVCC-HPPO) based on the improved Hierarchical Proximal Policy Optimisation (H-PPO) algorithm to jointly optimise traffic signal timing and Connected and Autonomous Vehicle (CAV) trajectories under mixed vehicular environments with both CAVs and Human-Driven Vehicles (HDVs). A hierarchical hybrid action space is designed to effectively constrain CAV acceleration and signal timing adjustments while explicitly accounting for car-following dynamics near intersections, enabling flexible exploration within physical limits. The hybrid actor-critic architecture facilitates simultaneous optimisation of discrete and continuous actions through parallel actors guided by a global critic, balancing optimization effectiveness with training stability. A multi-objective reward function simultaneously minimises vehicle delay and fuel consumption and maximises ride comfort. The core improvement involves a layered entropy regularisation strategy within the H-PPO algorithm, which separately manages discrete and continuous entropy to enhance exploration efficiency and stability across hybrid action dimensions. Real-world intersections evaluation results demonstrate that SVCC-HPPO significantly outperforms benchmark methods TRANSYT and DRL-based algorithms, achieving reductions of up to 46.3% in delay, 59.5% in queue length, and 52.9% in fuel consumption, alongside a 177.4% improvement in average speed. Performance gains are further enhanced with shorter optimisation intervals and higher CAV penetration rates.

Roadside sensing is an important part of intelligent traffic management systems (ITMSs) for collecting and processing information. In order to better assess and maintain the stability and safety of objects in traffic scenes, all types of basic information are required. This paper proposes a monocular vision-based object parameter measurement and geolocation method to address the problems of high cost and limited information dimension of traditional roadside sensors. Object detection and geometric transformation mapping are combined to achieve efficient estimation of key physical parameters with input of monocular images, and global navigation satellite system (GNSS) information is further incorporated to obtain geolocation of the target. In the method, after the key target is recognized by the neural network-based object detection algorithm, the pixel-level 2D image information is mapped to a series of 3D spaces based on the construction of a geometric model, which leads to further computation of various physical parameters, realizing multi-parameter estimation under one method. The method overcomes the dependence on fixed environments or known references and is highly applicable to non-cooperative scenes. The effectiveness of the method is shown via the experiments in multiple real scenes.