To improve road traffic safety, this study investigates the potential and feasibility of using electrocardiogram (ECG) signals for driver fatigue detection. A simulated driving experiment was designed to collect raw ECG data from participants, from which typical time-domain, frequency-domain, and non-linear features were extracted. A fatigue recognition model was then constructed using a support vector machine (SVM). Grid search combined with cross-validation was employed to optimize the model’s hyperparameters. The results demonstrated that the optimal classification performance was achieved when the penalty parameter C = 1 and the kernel parameter γ = 0.1. Under this configuration, further evaluation yielded classification accuracy, precision, recall, specificity, and F1-score of 84.9

Most of existing 3D reconstruction techniques for wire arc additive manufacturing (WAAM) rely on laser scanning, which is associated with high costs and limited to offline operation. For complex components such as lattice structures, internal geometries cannot be scanned due to external occlusion, resulting in incomplete CAD models. To address this critical gap, we propose a real-time, cost-effective online reconstruction solution specifically developed for cylindrical rods fabricated via WAAM. The geometric morphology of the molten pool is defined as a hemispherical-capped cylinder. First, the 2D contour of the molten pool is segmented from images using morphological techniques. Then, the 3D coordinates of the molten pool's centroid are calculated via stereo vision triangulation, and the 2D pixel coordinates of contour points are converted to 3D spatial coordinates using back-projection technology. Finally, spatial circle fitting is performed on the 3D contour points, and digital 3D model reconstruction is achieved through CAD lofting technology. Experimental results demonstrate strong geometric accuracy, with an average surface error of 0.31 mm and a volumetric intersection-over-union (IoU) of 0.87. Furthermore, the reconstructed 3D models enable realistic numerical simulations-validated by tensile tests where the simulated load-displacement curve aligns closely with experimental data.

Understanding how multimodal factors and individual differences shape drivers’ situation awareness (SA) is critical for safe takeover performance in autonomous driving. However, conventional SHAP-based attribution assumes feature independence, which is violated by strong covariances and latent causal dependencies among driving-related factors, leading to biased interpretations. To address this limitation, this study proposes a Causally Consistent SHAP Explanation Framework (CCSF) that incorporates causal constraints through key factor localization and hierarchical causal structure modeling. By adopting causal Shapley values, CCSF enables unified SA prediction and causally informed individual-level explanations. Using a high-performance LightGBM model, six key predictors were identified: frontal and occipital α/θ ratios, occipital theta relative power, NASA-TLX workload, pupil changes, and horizontal gaze dispersion. Guided by domain-constrained causal discovery, these features were organized into a theory-constrained causal structure, supporting a model-based interpretation that links SA formation primarily to workload-related neural rhythm modulation and visual sampling. Notably, pupil changes were more consistent with workload-related synchronous responses than with a direct pathway to SA. Under these constraints, consensus clustering identified three pathway-level attribution tendencies—neural-dominant, neural–visual combined, and visual-dominant—highlighting heterogeneity in model-based causal contributions. These findings provide a mechanistic foundation for improving system reliability and takeover safety in autonomous driving.

Conditionally automated driving (SAE Level 3) requires drivers to regain adequate situation awareness (SA) and resume control within a limited takeover window, making robust SA monitoring essential for safe deployment. However, existing SA prediction models often generalize poorly to unseen drivers due to substantial inter-subject variability in neurophysiological and visual responses. This study addresses cross-subject SA modeling by proposing a Statistical-Prior-Guided Deep Invariant Feature Attention Network (SP-DIFAN). The framework integrates: (i) feature-level statistical priors derived from effect-size analysis to guide a context-aware attention mechanism toward SA-discriminative cues; (ii) entropy-based sparsity regularization to promote concentrated and reproducible soft feature gating; and (iii) class-conditional subject-invariant regularization to mitigate crosssubject distribution shift while preserving SA-related class structure. A multi-task architecture jointly predicts binary SA sufficiency and continuous SAGAT scores, improving representation robustness across objectives. We conducted a high-immersion driving simulator study with 56 participants completing 18 takeover trials each. Synchronized EEG (256 Hz) and eye-tracking (50 Hz) signals were summarized into 59 trial-level features spanning takeover request onset to takeover initiation. Under rigorous leave-one-subject-out evaluation, SPDIFAN achieved strong cross-subject generalization (AUC = 0.824; Pearson's r = 0.771), yielding more balanced performance than representative linear, kernel, and ensemble baselines. Ablation results verified complementary benefits of statistical prior guidance and subject-invariant regularization. Interpretation further revealed stable cross-subject neural-ocular signatures coupling visual sampling dynamics, mirror-monitoring behavior, and EEG indicators of workload and vigilance. The proposed framework supports real-time SA assessment and informs adaptive takeover assistance and human-centered HMI design for Level 3 automation.

To address the problems of strong dependence on expert experience, low scheme generation efficiency, and difficulty in balancing multiple objectives in tunnel boring machine (TBM) configuration under complex geological conditions, this paper proposes a modular configuration method integrating improved Bayesian inference and multi-objective optimization. First, a generic bill of materials (GBOM) model for the TBM product family is established to describe module hierarchies, instance attributes, and configuration constraints. Second, the naive Bayes model is improved by combining mutual-information-based weighting and expert prior correction, so as to realize module-instance inference for each module family. Third, a three-objective optimization model is formulated with configuration cost, delivery time, and adaptability risk as the objectives and solved using an improved NSGA-II algorithm. Finally, the proposed method is validated using 27 historical engineering cases through leave-one-out cross-validation, supplementary sensitivity analysis, retrospective validation, and a synthetic scale-expansion experiment. The results show that the proposed framework achieves a mean LOOCV accuracy of 72.75% at the module-family level, exhibits robustness to moderate subjective-weight perturbations. A retrospective validation based on case indicates that the historical scheme is feasible but not globally balanced in the three-objective sense, while the proposed method can provide a clearer trade-off boundary for engineering decision-making. The study provides a feasible methodological reference for rapid TBM modular configuration under complex working conditions.

Due to various limitations such as technological capabilities, condi- tionally automated driving vehicles require driver intervention when they exceed the operational design domain (ODD) of their automated driving functions. The safety, stability, and comfort of the takeover process are collectively referred to as takeover performance, a topic that has garnered widespread attention among researchers. Based on an analysis of existing research on takeover performance, it was found that the evaluation methods for driver takeover performance are not comprehensive. Therefore, this study constructs a system of evaluation indicators for takeover performance based on multi-source data extraction of original evalu- ation indicators. This has significant implications for optimizing takeover request strategies in autonomous driving scenarios. The specific research content includes: (1) Designing and conducting takeover experiments based on driving simulators to simulate the takeover process of automated driving under different driver states and takeover request times; (2) Extracting original evaluation indicators of driver takeover process from multi-source data, including subjective situational awareness, subjective task load, takeover responsiveness, takeover effective response time, takeover stability, and takeover safety; (3) Establishing a system of evaluation indicators for takeover performance based on factor analysis and analyzing the differences in driver takeover performance from multiple dimensions under the influence of different factors.

Driver disengagement, including operational and attentional disengagement, significantly increases driving risk during human-vehicle co-driving, necessitating effective monitoring of disengagement behaviors. Unlike existing studies that rely on third-person perspective (TPP) visual information for indirect inference, this study proposes a wearable sensing-based driver disengagement recognition system, termed HV4DDR, which enables fine-grained recognition through the fusion of hand-motion sensing and first-person perspective (FPP) visual information. The system comprises a flexible smart glove, smart glasses, and a deep learning-based recognition algorithm. Specifically, a self-powered piezoelectric nanogenerator-based sensing unit (P-SU) is developed and integrated into a flexible smart glove (P-FSG) to directly capture temporal behavioral features of the driver’s hands, a primary site of disengagement, while smart glasses synchronously record first-person visual attention. To effectively exploit the multimodal data, a knowledge distillation network (KITE) is constructed to learn the intrinsic mapping between sensing signals and disengagement behaviors, enabling lightweight yet high-performance inference suitable for vehicular edge computing scenarios. Experimental results on a wearable multimodal dataset for driver disengagement recognition (WM-DDR) demonstrate an accuracy of 99.92% with only 3.29 million parameters and an inference latency of 10.41 ms. Furthermore, real-vehicle and simulation-based experiments validate that the developed DriCare system based on HV4DDR significantly enhances driving safety under both manual and automated driving modes. This work establishes a scalable and practical framework for real-time, fine-grained driver disengagement recognition, highlighting the potential of self-powered piezoelectric nanogenerator-enabled wearable sensing systems for intelligent transportation safety.

With the rapid development of information and network technologies, intelligent connectivity has become a dominant trend in the automotive industry. While providing a richer driving experience for its users, intelligent connected vehicles bring about more complex challenges to driving safety. The touch screens and diverse human-machine interaction methods of intelligent connected vehicles increase the possibility of visual distractions for drivers. Therefore, research on visual distraction in the context of intelligent connected vehicles (ICVs) holds substantial practical importance. Based on the human-machine interaction methods characteristics of intelligent connected vehicles, this paper designed and conducted a visual distraction driving simulation experiment, collecting data on driving performance, eye movement, and subjective workload evaluation data. Firstly, this paper designed a visual distraction driving sub-task and collected driving performance data, including vehicle speed, acceleration, and steering wheel angle speed, through a driving simulator. Simultaneously, the Tobii Glass 2 eye tracker was used to record pupil diameter data, and the NASA-TLX and SWAT subjective workload scales were employed to assess the driver's mental load. The driving performance, eye movement, and subjective workload characteristics of drivers with different driving experiences were analyzed when performing distraction tasks of varying difficulty. Finally, through one-way ANOVA and Pearson correlation coefficient analysis, five key parameters were identified as visual distraction indicators: longitudinal speed standard deviation, longitudinal acceleration standard deviation, lateral acceleration standard deviation, steering wheel angular velocity, and pupil diameter. The differences in control stability among drivers with different demographic characteristics under distracted conditions were also compared.

In high-level autonomous driving, driver uncertainty during lane-changing takeovers presents a major challenge to human-machine coordination and traffic safety. This study explores driver physiological and behavioral responses under uncertainty, establishing a multidimensional evaluation framework. A driving simulator experiment is designed with controlled variables including time-to-collision (TTC), inter-vehicle gap, and relative speed. Physiological signals—ECG, EDA, and EMG—alongside driving behavior data are collected and time-aligned across takeover phases. Statistical analysis shows that EDA and EMG metrics significantly differ across uncertainty levels, while reaction time, lateral velocity, steering angle, and lane deviation also vary accordingly. Reaction time is negatively correlated with subjective uncertainty ratings. Moreover, uncertainty peaks are delayed with increasing relative speed but show no linear relationship with TTC or gap. These findings confirm the feasibility and effectiveness of integrating physiological and behavioral data for uncertainty assessment in autonomous takeover scenarios.

Monocular depth estimation for UAV cargo delivery must balance limited onboard computational resources against high accuracy requirements in safety-critical near-field regions (0-30 meters). Inspired by human myopic vision principles, this work presents HMV-Net, a bio-inspired framework that reallocates computational resources toward operationally critical foreground regions without introducing inference overhead. The framework implements two key components: a training-time attention mechanism that imposes no deployment cost, and a parameter-free refinement module for identified critical regions. Compared with uniform processing baselines, HMV-Net improves depth accuracy by 1.6%, reduces RMSE by 2.1%, and enhances edge accuracy from 0.512 to 0.562, while maintaining identical parameter count (24.79M) and throughput (187 vs 192 FPS on desktop GPU). Embedded deployment on Jetson Orin NX achieves over 33 FPS with 1.8-5.0% accuracy improvements across NYU-v2, KITTI, and COCO2017, demonstrating its potential for resource-constrained UAV platforms. Code is available at https://github.com/tian1bukewei/HMV-Net.

In the foreseeable future, human drivers will remain an indispensable role of autonomous vehicles, rendering real-time and accurate recognition of driver operational behaviors critical for driving safety and cooperative decision-making. Although flexible tactile sensors offer a promising solution for capturing driver-vehicle interactions, driver operational behaviors span a wide range of interaction forces, from subtle pre-operational cues to large-amplitude control actions, posing a significant challenge for conventional tactile sensors to achieve both high sensitivity and a broad sensing range. Inspired by the hierarchical mechanosensing mechanism of human skin, this study proposes a biomimetic gradient-structured flexible magnetic skin (BG-FMS) for fine-grained driver operational behavior recognition. The designed three-gradient elastic modulus architecture enables the BG-FMS to simultaneously provide high sensitivity to subtle operational cues and a wide dynamic range for large-amplitude driving actions, while maintaining fast response and strong environmental robustness. The BG-FMS is seamlessly integrated on the steering wheel and pedal surfaces, enabling real-time acquisition of driver operational signals. Meanwhile, a channel attention-enhanced deep learning model (CA-TCNet) is developed to accurately recognize driver operational behaviors. Experimental results demonstrate that the proposed driver operational behavior recognition system achieves recognition accuracies of 99.6%, 98.0%, and 90.7% using 3s, 2s, and 1s of operation data, respectively, highlighting its ultra-fast early-stage behavior recognition capability. This work presents a robust, non-intrusive, and deployable human-vehicle interaction sensing solution, thereby broadening the application scope of flexible magnetic skins in intelligent transportation and cooperative driving systems.

Autonomous vehicles (AVs) provide an effective solution for enhancing traffic safety. In the last few years, there have been significant efforts and progress in the development of AVs. However, the public acceptance has not fully kept up with technological advancements. Public acceptance can restrict the growth of AVs. This study focuses on investigating the acceptance and takeover behavior of drivers when interacting with AVs of different styles in various scenarios. Manual and autonomous driving experiments were designed based on the driving simulation platform. To avoid subjective bias, principal component analysis (PCA) and the Gaussian mixture model (GMM) were used to classify driving styles. A total of 34 young participants (male-dominated) were recruited for this study. And they were classified into three driving styles (aggressive, moderate, and conservative). And AV styles were designed into three corresponding categories according to the different driving behavior characteristics. This study reveals that drivers generally prefer driving scenarios with lower risk levels. When drivers perceive safety, they are more likely to adopt more efficient AVs. Additionally, drivers tend to accept AVs that align better with their driving styles. However, it is not found that more aggressive or conservative AVs have a significant impact on their acceptance. Takeover behavior has been identified as a significant mediator of acceptance, with the potential to influence drivers’ perceptions and attitudes. There is a marked decline in acceptance when takeover behavior happens. The results show that regulating takeover behavior is essential for the development of AVs that promote greater acceptance. And this study contributes theoretical support to the development of adaptive AVs.

In conditionally automated driving, driver capability and task demands are crucial for safe takeover transitions. Identifying factors influencing task demand and driver capability, as well as exploring their combined effects on driver behaviour and perception, are essential for developing models that optimise driver performance. To simulate varying task demands, we adjusted the urgency of the takeover time budget (TOTB) and the complexity of traffic scenarios (i.e., TOR-Lane, the lane where the vehicle was located when the takeover request occurred), while manipulating driver capability by introducing non-driving related tasks (NDRTs). A multilevel modelling approach was employed to analyse how these factors jointly influenced takeover behaviour and situation awareness (SA). Results indicated that TOTB, NDRT, and TOR-Lane influenced takeover timeliness at different time stages: NDRT affected driver reaction time, while TOTB and TOR-Lane impacted information processing time (IPT). A shorter TOTB resulted in reduced IPT and lower minimum time-to-collision [min (TTC)], especially when visual-cognitive NDRT were involved, which further impaired takeover quality. Moreover, increased traffic environment complexity prolonged IPT and reduced min (TTC). To meet task demands, drivers adjusted their visual behaviour to rapidly restore SA by reducing the quality of visual processing for low-priority elements, thereby prioritising resources to takeover tasks. Participants' SA improved as TOTB increased, reaching saturation levels that varied with scenario complexity-7-9 s in the centre lane and 5-7 s in the side lane. This study reveals how driver behavioural patterns are influenced by task demands and their own capabilities, supporting the design of adaptable human-machine interaction models.

Driver trust in Automated Driving Systems (ADS) is a key factor for ensuring human-vehicle-cooperative driving safety. This study focuses on this aspect and conducts driving simulation experiments on a static driving platform. Under various scenarios involving different driving takeover events, vehicle styles, and takeover warning types, the study uses driver eye-tracking data and vehicle status data to predict driver trust. An Attention-CNN model, combining multi-scale convolution and attention mechanisms, is employed for trust prediction, and Shapley values are used to determine the importance of each feature to optimize the model. The experimental results show that the model performs well in predicting driver trust, with an accuracy of 80.6% and an F1 score of 81.4%, representing a significant improvement over the baseline model. This provides effective methodological support for driver trust evaluation in Automated Driving Systems.

The takeover issue, especially the setting of the takeover time budget, is a critical factor restricting the implementation and development of conditionally automated vehicles. The general fixed takeover time budget has certain limitations, as it does not take into account the driver's non-driving behaviors. Here, we propose an intelligent takeover assistance system consisting of all-round sensing gloves, a non-driving behavior identification module, and a takeover time budget determination module. All-round sensing gloves based on triboelectric sensors seamlessly detect delicate motions of hands and interactions between hands and other objects, and then transfer the electrical signals to the non-driving behavior identification module, which achieves an accuracy of 94.72% for six non-driving behaviors. Finally, combining the identification result and its corresponding minimum takeover time budget obtained through the takeover time budget determination module, our system dynamically adjusts the takeover time budget based on the driver's current non-driving behavior, significantly improving takeover performance in terms of safety and stability. Our work presents a potential value in the application and implementation of conditionally automated vehicles.

In mixed traffic environments with autonomous and traditional vehicles, perceiving the lane change intention of vehicles in advance is crucial to ensure traffic safety. Considering the current issue of lane change intention detection methods overemphasize the surrounding features of the host vehicle and have low accuracy, this paper proposes a lane change intention prediction algorithm that only fuses the vehicle forward features. This study first extract lane-changing trajectory data of vehicles that meet the definition of lane change from the NGSIM dataset based on changes in the lane marking. Then, the motion features and the forward traffic state features of the host vehicle are extracted from these lane-changing trajectory data to construct the dataset for predicting the vehicle’s lane change intention. Finally, we use the LSTM_Self-Attention model to predict the lane-changing intention for different lead times. The results show that the LSTM_self-Attention model proposed in this study performs well for predicting vehicle’s lane change intention. The prediction accuracy can reach 82

The superposition effect of various cutting mechanisms (CM) in the fine drilling process brings great challenges to the accurate characterization of the cutting stress field of the workpiece. To solve the above problem, the cutting stress characterization modeling and parameter identification for the fine drilling process with multiple cutting mechanisms is studied in this paper. Firstly, two cutting mechanisms (shear-slip and plough-slip) are distinguished according to the relative tool sharpness (RTS) which is determined by the cutting tool radius and cutting depth, and the fine characterization model for drilling stress of the workpiece is constructed by considering the two cutting mechanisms. Then, in order to overcome the problem that model parameters are difficult to be accurately determined, the sub-interval decomposition optimization method (SDOM) and the improved particle swarm optimization (PSO) are employed to identify parameters in the model. Finally, the proposed method is verified by comparing the single cutting mechanism model, the multiple cutting mechanisms model, and the actual characterization parameter model.

The provision of real-time, accurate perception of vulnerable road users (VRUs) via infrastructure-sensors-based devices is integral to roadside perception in vehicle-infrastructure collaboration system. However, prevailing data and algorithms fall short of accomplishing this task effectively on high-resolution imagery. In response, we introduce a visual perception framework, VRUFinder, designed specifically for infrastructure-enabled deployment, and a multi-view symmetrical knowledge distillation methodology for VRU recognition. This approach amalgamates various teacher networks into streamlined student networks from diverse perspectives. By integrating our novel logical connectivity and quality judgment model, we enhance the existing state-of-the-art algorithms of YOLOv7 and StrongSORT. Moreover, we present VRUNet, a novel dataset for VRU recognition, furnishing high-resolution, top-down perspective images with visual sensor acquisition system. To the best of our knowledge, datasets of this nature are seldom found in current VRU recognition research. The effectiveness of our approach is substantiated through a series of ablation experiments and engineering case study on a low computational infrastructure-sensor-enabled device. By encapsulating our approach, we provide mature solutions for commercial infrastructure-sensors-based devices, which will contribute to the development of connected and automated vehicles and intelligent transportation systems.

For autonomous vehicles, real-time and accurate longitudinal driving intention recognition is crucial as it effectively enhances driving safety and improves the driving experience. This study proposes a novel data and model hybrid-driven fine-grained longitudinal driving intention prior recognition system (LDIPRS). Firstly, the system integrates a human-pedal interaction sensor (HPIS) based on triboelectric nanogenerators for fine-grained longitudinal driving maneuver monitoring and the channel attention (CA)-enhanced convolutional neural network (CBRCNet). The HPIS, integrated into the vehicle's acceleration and brake pedals, is capable of monitoring driver foot movement information in the form of electrical signals before the vehicle responds, achieving data level advance. The collected electrical signals are fed into the CBRCNet network, which models and learns the mapping relationship between these signals and fine-grained longitudinal driving intentions, leading to model level advances. The HPIS completes the capture of longitudinal maneuver information 541 ms before the driving simulator starts to respond at the data level. At the model level, CBRCNet can achieve a recognition accuracy of 96.1 % based on partial response data (50 ms after starting response) rather than complete response data of the HPIS. Finally, our proposed LDIPRS realizes the recognition of emergency braking, rapid acceleration, normal braking, and normal acceleration in advance by 732 ms, 1035 ms, 1757 ms, and 2227 ms, respectively. This study introduces self-powered, low-cost, highly sensitive triboelectric sensors into the field of intention recognition, and combines the triboelectric sensors with deep learning algorithms to offer a promising solution to improve the safety of autonomous vehicles and the efficiency of intelligent transportation systems.

开发了可伸缩的聚丙烯酰胺(PAAM)-licl基摩擦纳米发电机(PL-TENG)和具有铝(AI)-Kapton摩擦层结构的摩擦纳米发电机(AK-TENG),用于驾驶人的驾驶行为监测.PL-TENG具有体积小、质量轻、弹性好、透明度高、生物相容性好、无创、耐用、生产简单、成本低等优点.将PL-TENG贴在人体皮肤表面不会影响人体的舒适度,可在不干扰正常驾驶的情况下,获取驾驶人头部和面部的微小动作信息.AK-TENG具有高灵敏度和一定的柔性,相比于普通摩擦纳米发电机的结构更加坚固稳定且不易损坏,用于获取驾驶人在转向盘和踏板上的驾驶操作行为特征.本文利用PL-TENG和AK-TENG获取驾驶行为信息数据,获取驾驶人的驾驶状态和操作行为特征,同时也为智能交通系统中高灵敏度自供电传感器的设计提供了新的思路.