The increasing prevalence of roadworks poses significant challenges to maintaining accurate and up-to-date high-definition (HD) maps, crucial for autonomous vehicle (AV) safety and efficiency. Current methods for updating HD maps are expensive, time-consuming, and not responsive to real-time changes. This paper proposes a real-time, low-cost pipeline for updating HD maps with roadworks information using a monocular camera and a Contrastive Language–Image Pre-Training-based Vision Language Model (VLM), achieving robust few-shot sign recognition with minimal annotated data. The workflow is designed for low-cost, real-time deployment using only monocular camera input, and supports rapid, incremental HD map updates directly in OpenDRIVE format. Extensive experiments demonstrate that our system outperforms conventional baselines (e.g., fine-tuned You Only Look Once (YOLO) v11) not only in data-scarce settings but also across challenging environmental conditions. The recognition model is trained on a diverse dataset of 3752 real and virtual images, enhanced through data augmentation techniques. The model achieves a 97.12% recognition rate on a test dataset of 752 images and a root mean square error (RMSE) of less than 1.2 m for positional accuracy, processing single image inputs in 1.54 s. By leveraging the OpenDRIVE format, this approach ensures seamless data exchange between different HD map systems, facilitating real-time updates that accurately reflect current road conditions. The methodology demonstrates significant benefits in terms of responsiveness, cost and time efficiency, enhanced safety, and flexibility. Trials on the United Kingdom motorways validate the pipeline's effectiveness, offering a robust solution to dynamic road conditions and enabling safer, more efficient AV navigation.

As autonomous vehicles (AVs) advance, the integration of Large (Vision) Language Models (LLMs and VLMs) has emerged as a promising approach to enhance AV capabilities in perception, planning, decision-making, and data generation. However, the practical challenges of incorporating LLMs and VLMs into AV systems, including computational efficiency, real-time processing, and ethical considerations, remain underexplored. This survey aims to provide a comprehensive review of the current research on LLM and VLM applications in AVs, focusing on the following key areas: modular integration, end-to-end integration, data generation, evaluation platforms, datasets, and benchmarks. We systematically analyse 77 recent papers published before Sep 2025, detailing their methodologies and models. Our findings highlight the potential of LLMs and VLMs to improve AV system performance while acknowledging limitations. This survey offers researchers and practitioners a panoptic view of the classification and progression of LLMs and VLMs in the AV sphere, while systematically distilling models to their core components. We envision this survey as a central reference for AV researchers navigating this rapidly evolving landscape to accelerate future research.

The efficiency of urban traffic management and congestion alleviation relies heavily on accurate forecasting of Origin-Destination (O-D) demand matrices. Existing models primarily focus on estimating O-D demand for various travel purposes throughout the day, which is characterised by its pulsating nature. However, these models often compromise the precision of peak-hour forecasts, leading to unreliable dynamic traffic control and challenges in effectively reducing peak-hour congestion. To tackle this challenge, this paper proposes a novel method for predicting commuting O-D demand matrices. Our method employs community detection algorithms on road networks to precisely partition commute O-D regions, incorporating Points of Interest (POIs). We also present a spatio-temporal dynamic weighted hypergraph model that leverages these partitioned regions, time characteristics from observed O-D trips, and meteorological data to improve forecasting. Comparative analyses with contemporary models and ablation studies indicate our method significantly enhances prediction accuracy, by approximately 5%. These findings imply that the proposed method more effectively encompasses the varied characteristics of commuting during peak hours, thereby providing more accurate demand matrices for urban traffic management.

Navigating through roadworks represents one of the main sources of safety risk for Connected and Autonomous Vehicles (CAVs) due to the altered road layouts. The built-in base maps do not normally reflect these changes, causing CAVs to experience difficulties in sensing and trajectory generation. Therefore, the objective of this paper is to evaluate different collision-free trajectory generation for CAVs at roadworks to improve safety and traffic performance. Trajectory generation algorithms using lane-level dynamic maps were examined for: 1) CAVs rely on data from in-vehicle sensor only; and 2) CAVs receive additional information via a Smart Traffic Cone (STC) in advance regarding roadwork configurations. Experiments were conducted at a controlled motorway facility operated by National Highways (England) using a vehicle instrumented with a suite of sensors. Schematics of the roadworks scenario were translated into an integrated simulation platform consisting of a traffic microsimulation (VISSIM) to simulate traffic dynamics and a sub-microscopic simulator (PreScan) capable of simulating vehicle autonomy and connectivity. Results indicate that traffic conflicts and delays decrease by 40% and 3% respectively when CAVs receive additional information in advance (i.e., Scenario 2) compared to the other scenario. These findings would assist road network operators in developing ‘CAV-enabled roadworks’ and vehicle manufacturers in designing a vehicle-based ‘roadworks assist’ system.

Collision warning systems (CWSs) have been recognized as effective tools in preventing vehicle collisions. Existing systems mainly provide safety warnings based on single-directional approaches, such as rear-end, lateral, and forward collision warnings. Such systems cannot provide omni-directorial enhancements on driver’s perception. Meanwhile, due to the unclear and overlapped activation areas of above single-directional CWSs, multiple kinds of warnings may be triggered mistakenly for a collision. The multi-triggering may confuse drivers about the position of dangerous targets. To this end, this paper develops a spatial-state-based omni-directional collision warning system (S-OCWS), aiming to help drivers identify the specific danger by providing the unique warning. First, the operational domains of rear-end, lateral, and forward collisions are theoretically distinguished. This distinction is attained by a geometric approach with a rigorous mathematical derivation, based on the spatial states and the relative motion states of itself and the target vehicle in real time. Then, a theoretical omni-directional collision warning model is established using time-to-collision (TTC) to clarify activation conditions for different collision warnings. Finally, the effectiveness of the S-OCWS is validated in field tests. Results indicate that the S-OCWS can help drivers quickly and properly respond to the warnings without compromising their control over lateral offsets. In particular, the probability of drivers giving proper responses to FCW doubles when the S-OCWS is on, compared to when the system is off. In addition, the S-OCWS shortens the responses time of nonprofessional drivers, and therefore enhances their safety in driving.

Existing autonomous driving systems of intelligent vehicles such as advanced driver assistant systems (ADAS) assess and quantify the level of potential safety threats. However, they may not be able to plan the best response to unexpected dangerous situations and do not have the ability to cope with uncertainties since not all vehicles can always keep a safe gap from preceding vehicles and drive at a desired velocity. Previous research has not taken such uncertainties into account, it is, therefore, necessary to develop models which are not restricted by the predefined movement patterns of a vehicle. Existing systems are based on a model that estimates the threat level based only on one factor Time-To-Collision (TTC). This approach is limited since it cannot handle all scenarios and ignores all uncertainties. To overcome these limitations, this paper utilised deep learning to develop a range of models that rely on a group of factors to reliably estimate the threat level and predict conflicts under uncertainty using the concept of looming ’. Comparative analyses were undertaken by incorporating new varying input factors to each model (e.g., surrogate safety measures, vehicle kinematics, macroscopic traffic data). Real-world experiments demonstrated that adding new factors increases the reliability and sensitivity of the models. Results also indicated that the models that consider looming provide low false alarm rate extending their applications for a wider spectrum of traffic scenarios. This is paramount for ADAS as uncertainties are inherent in the deployment of connected and autonomous vehicles in a mixed traffic stream.

Real-time crash risk prediction models aim to identify pre-crash conditions as part of active traffic safety management. However, traditional models which were mainly developed through matched case-control sampling have been criticised due to their biased estimations. In this study, the state-of-art class balancing method known as the Wasserstein Generative Adversarial Network (WGAN) was introduced to address the class imbalance problem in the model development. An extremely imbalanced dataset consisted of 257 crashes and over 10 million non-crash cases from M1 Motorway in United Kingdom for 2017 was then utilized to evaluate the proposed method. The real-time crash prediction model was developed by employing Deep Neural Network (DNN) and Logistic Regression (LR). Crash predictions were performed under different crash to non-crash ratios where synthetic crashes were generated by Wasserstein Generative Adversarial Network (WGAN), Synthetic Minority Over-sampling Technique (SMOTE) and Adaptive Synthetic (ADASYN) sampling respectively. Comparisons were then made with algorithmic-level class balancing methods such as cost-sensitive learning and ensemble methods. Our findings suggest that WGAN clearly outperforms other oversampling methods in terms of handling the extremely imbalanced sample and the DNN model subsequently produces a crash prediction sensitivity of about 70% with a 5% false alarm rate. Based on the findings of this study, proactive traffic management strategies including Variable Speed Limit (VSL) and Dynamic Messing Signs (DMS) could be deployed to reduce the probability of crash occurrence.

Current approaches to estimate the probability of a traffic collision occurring in real-time primarily depend on comparing traffic conditions just prior to collisions with normal traffic conditions. Most studies acquire pre-collision traffic conditions by matching the collision time in the national crash database with the time in the traffic database. Since the reported collision time sometimes differs from the actual time, the matching method may result in traffic conditions not representative of pre-collision traffic dynamics. In this paper, this is overcome through the use of highly disaggregated vehicle-based traffic data from a traffic micro-simulation (i.e., VISSIM) and the corresponding traffic conflicts data generated by the surrogate safety assessment model (SSAM). In particular, the idea is to use traffic conflicts as surrogate measures of traffic safety so that traffic collisions data are not needed. Three classifiers (i.e., support vector machines, k-nearest neighbours, and random forests) are then employed to examine the proposed idea. Substantial efforts are devoted to making the traffic simulation as representative of the real-world as possible by employing data from a motorway section in England. Four temporally aggregated traffic datasets (i.e., 30 s, 1 min, 3 min, and 5 min) are examined. The main results demonstrate the viability of using traffic micro-simulation along with the SSAM for real-time conflicts prediction and the superiority of random forests with 5-min temporal aggregation in the classification results. However, attention should be given to the calibration and validation of the simulation software so as to acquire more realistic traffic data, resulting in more effective prediction of conflicts.

The concept of user-level integrity monitoring has been successfully applied to air transport navigation systems, where the main focus is on the errors associated with the Global Positioning System (GPS)-data-processing chain. Little research effort has been devoted to the study of integrity monitoring for the case of land vehicle navigation systems. The primary difference is that it is also necessary to consider errors associated with a spatial map and a map-matching (MM) process when monitoring the integrity of a land vehicle navigation system. This is because these two components play a vital role in land vehicle navigation. To date, research has focused on either the integrity of raw positioning data obtained from GPS or the integrity of the MM process and digital map errors. In this paper, these sources of error are simultaneously considered. Therefore, the main contribution of this paper is to report the development of a user-level integrity-monitoring system that concurrently takes into account all the potential error sources associated with a navigation system and considers the operational environment to further improve performance. Errors associated with a spatial road map are given special attention. Two knowledge-based fuzzy inference systems were developed to measure the integrity scale. The performance of the integrity method was assessed using field data collected in Nottingham and London, U.K. The results indicate that the integrity method provides valid warnings 98.2% and 99.4% of the time for positioning data in a mixed operational environment in Nottingham and suburban areas of London, respectively.