Autonomous vehicles (AVs) will coexist with human-driven vehicles (HVs) for a long transition period. During this time, traffic safety may be compromised as various styles of AVs increase the uncertainty of surrounding human drivers’ risk perception and decision-making. This study conducted a cooperative driving simulator experiment to investigate the impact of different AV styles on drivers’ risk perception in overtaking situations. Thirty-two drivers participated, experiencing varying levels of AV aggressiveness and velocity. Verbal protocol analysis assessed drivers’ risk perceptions and risk avoidance decisions. SHapley Additive exPlanations were utilized to identify the factors influencing subjective and objective risk levels. The results indicated that subjective risk levels are primarily influenced by relative movements, while objective risk levels are predominantly affected by the movements of AVs. This study lays the groundwork for improving AV trajectory planning and assisting human drivers in adapting to interactions with AVs.

In zero-shot setting, test-time adaptation adjusts pre-trained models using unlabeled data from the test phase to enhance performance on unknown test distributions. Existing cache-enhanced TTA methods rely on a low-entropy criterion to select samples for prototype construction, assuming intra-class compactness. However, low-entropy samples may be unreliable under distribution shifts, and the resulting prototypes may not ensure compact intra-class distributions. This study identifies a positive correlation between cache-enhanced performance and intra-class compactness. Based on this observation, we propose a Multi-Cache enhanced Prototype-based Test-Time Adaptation (MCP) featuring three caches: an entropy cache for initializing prototype representations with low-entropy samples, an align cache for integrating visual and textual information to achieve compact intra-class distributions, and a negative cache for prediction calibration using high-entropy samples. We further developed MCP++, a framework incorporating cross-modal prototype alignment and residual learning, introducing prototype residual fine-tuning. Comparative and ablation experiments across 15 downstream tasks demonstrate that the proposed method and framework achieve state-of-the-art generalization performance. Project Page available at: https://zhaihaotian.github.io/MCP-ICCV25/

Early risk diagnosis and driving anomaly detection from vehicle stream are of great benefits in a range of advanced solutions towards Smart Road and crash prevention, although there are intrinsic challenges, especially lack of ground truth, definition of multiple risk exposures. This study proposes a domain-specific automatic clustering (termed AutoCluster) to self-learn the optimal models for unsupervised risk assessment, which integrates key steps of clustering into an auto-optimisable pipeline, including feature and algorithm selection, hyperparameter auto-tuning. Firstly, based on surrogate conflict measures, a series of risk indicator features are constructed to represent temporal-spatial and kinematical risk exposures. Then, we develop an unsupervised feature selection method to identify the useful features by elimination-based model reliance importance (EMRI). Secondly, we propose balanced Silhouette Index (bSI) to evaluate the internal quality of imbalanced clustering. A loss function is designed that considers the clustering performance in terms of internal quality, inter-cluster variation, and model stability. Thirdly, based on Bayesian optimisation, the algorithm auto-selection and hyperparameter auto-tuning are self-learned to generate the best clustering results. Herein, NGSIM vehicle trajectory data is used for test-bedding. Findings show that AutoCluster is reliable and promising to diagnose multiple distinct risk levels inherent to generalised driving behaviour. We also delve into risk clustering, such as, algorithms heterogeneity, Silhouette analysis, hierarchical clustering flows, etc. Meanwhile, the AutoCluster is also a method for unsupervised data labelling and indicator threshold calibration. Furthermore, AutoCluster is useful to tackle the challenges in imbalanced clustering without ground truth or a priori knowledge.

Road accidents wreck lives. Could technology stop them from happening? Driving better road safety with technology and artificial intelligence are the key elements considered by several carmakers. The key aspect of transportation in the future is to build an ecosystem comprising autonomous, connected, electric and shared mobility. The evolution of autonomous vehicles (AVs) can potentially aid transportation to people and be deployed to resolve mobility-related pain for drivers and safety on roads while changing lanes. Thus, the intelligent assistance system should be smart enough to track such vehicles while deviating into another lane. In this paper, we propose a lane change prediction framework for feature learning, with the aim to have a deep and comprehensive understanding of lane change behaviors, meanwhile, reach a high performance based on the selected features. A time-step dataset with more than 1000 features is constructed from vehicle trajectory data. To identify the key features involved in the original feature set, an XGBoost-based three-step feature learning algorithm is proposed, which integrates the feature importance ranking, metric selection and recursive feature elimination. After analyzing the accuracy of test data from different time segment positions, the sliding window method is applied on a time-step dataset with filtered features to properly select time segments, which are flattened into corresponding time-series dataset for model prediction. In our case studies, a publicly available dataset, Next Generation SIMulation (NGSIM), is adopted to conduct experiments of feature learning and lane change prediction, where we achieved a new state-of-art accuracy with 97.6% under the time-series data of 75 selected features and 1-second window size with predictor XGBoost after adopting proposed three-step method, which is superior to the other state-of-the-art feature selection methods.

This study presents a domain-specific automated machine learning (AutoML) for risk prediction and behaviour assessment, which can be used in the behavioural decision-making and motion trajectory planning of autonomous vehicles (AVs). The AutoML enables end-to-end machine learning from vehicle movement and sensing data to detailed risk levels and corresponding behaviour characteristics, which integrates three main components of: unsupervised risk identification by surrogate risk indicators and big data clustering, feature learning based on XGBoost, and model auto-tuning by Bayesian optimisation. Then, the functions and performance of AutoML are evaluated based on NGSIM data, with assumptions of various sensing configurations or data acquisition conditions. AutoML achieves satisfactory results of behaviour-based risk prediction, which has a predictive power of 91.7% overall accuracy for four risk levels, and about 95% accuracy for safe-risk distinction. Bayesian optimisation guides the self-learning of AutoML to get the optimised feature subsets and hyperparameter values. The identification of key features not only produces better performance with fewer computation costs, but also provides data-driven insights about AV design, such as sensor configurations and sensor data mining, from risk decision-making perspectives. The application potentials of AutoML in AVs are discussed.