Yongjun Shen

The mixture of autonomous vehicles (AVs) and human-driven vehicles (HVs) dominates the current road traffic. AVs can record their own motion and measure the trajectories of surrounding HVs, and these recorded trajectories support mixed-driving research. Due to differences in measurement processes and complex traffic environments, HV trajectories generally exhibit stronger local perturbations than AV trajectories. However, existing mixed-driving research rarely accounts for this difference in local disturbances, which may affect subsequent analyses. This study proposes a local disturbance alignment (LDA) method to align local perturbations in HV and AV trajectories by optimizing filtering parameters. The effectiveness of LDA is validated through car-following modeling and risk prediction on Argoverse 2 and waymo open motion dataset (WOMD). Meanwhile, the differences among car-following pairs are revealed through the proposed resilience measurement framework. The results show that: (1) The differences in local disturbances between HV and AV trajectories are widespread in mixed-driving trajectory datasets processed by various filtering approaches; (2) The proposed LDA substantially reduces errors in car-following modeling and risk prediction compared with the classical filtering approach; (3) AVs exhibit a favorable car-following pattern characterized by assertive acceleration and gentle deceleration, and they achieve the highest and most stable car-following resilience scores. This study provides a novel perspective on mixed-driving trajectory processing and safety evaluation, and the processed WOMD car-following samples are publicly released ( https://github.com/HuansongZhang/Waymo-Open-Motion-Data-Car-following ).

Lane-change intention prediction is critical for intelligent vehicles, and driver decisions depend on the perception and processing of driving context information. Despite advances in deep learning in this domain, further exploration of the driving context processing remains essential. This study proposes a soft mask attention mechanism to adaptively enhance or suppress input features from target and surrounding vehicles. After that, the masked features are extracted using a heterogeneous information encoder, thereby differentiating between target and surrounding vehicle information processing. The encoded features are then integrated using a multi-head attention mechanism, and the lane change probabilities are output through convolution operations and decoder layers. Experiments demonstrate: (1) The proposed SMILE-LC (Soft Mask Information with Lane-based Encoding for Lane Change) achieves optimal prediction performance across all perception ranges with strong efficiency and generalization. (2) The soft mask attention mechanism intuitively reveals the information processing patterns in prediction model. The target vehicle’s lateral acceleration is the most critical feature, while the position information of surrounding vehicles is more important than their velocity and acceleration. (3) The heterogeneous information encoder significantly improves lane change prediction performance, and the proposed lane-based encoding strategy outperforms other encoding architectures. These results can significantly advance the development of Advanced Driver Assistance Systems (ADAS) and enhance lane-change safety.

Although autonomous vehicles leverage perceptual information to execute downstream tasks such as prediction and control to enhance driving safety, they are vulnerable to cyberattacks. Attackers can manipulate model outputs by tampering with inputs. This study proposes a Counterfactual Generation Trend Attack (CGTA) framework for predicting rear-end collision risk. It manipulates inputs within the Key Feature Region using an optimization algorithm, aiming to either elevate (increase attack) or degrade (decrease attack) the predicted driving safety. Furthermore, considering scenarios involving partial knowledge stealing where model parameters are inaccessible, an attack scheme based on model distillation is proposed. After processing trajectory data and quantifying rear-end collision risk, experiments were conducted on three mainstream prediction models. The results demonstrate that: (1) CGTA can effectively execute real time targeted attacks. Simultaneously, the resistance to attack varies under attack directions and models. The deviation of model outputs from ground truth was from 2.87% to 54.75% under increased attacks and from −3.29% to −86.62% under decreased attacks; (2) the multi-head attention (MHA) exhibited superior attack resistance. The attack effects of the prediction models reduced by 41.69% to 86.10% after incorporating MHA; and (3) although the attack effectiveness under partial knowledge stealing decreased, it realized the target attack trend. In addition, this study provides a detailed sensitivity analysis of the parameters and verifies the generalization of the CGTA. The findings reveal the feasibility and effect mechanisms of targeted cyberattacks on autonomous driving systems, while offering theoretical support for defense against such attacks.

As autonomous-driving technology advances, autonomous vehicles (AVs) and human-driven vehicles (HVs) are expected to coexist for an extended period. While resilience is widely applied in macro safety analysis, its application in micro-process modeling remains underdeveloped. This study proposes a resilience-based framework comprising three phases (initial, safety decay, safety recovery) to investigate the evolution of driving risk across three car-following scenarios: HV–HV (HV following HV), HV–AV (HV following AV), and AV–HV (AV following HV). The analytical framework consists of three steps: First, car-following pairs with complete risk-evolution processes were identified. Second, cluster analysis was applied to categorize phase-specific patterns. Third, the association rule mining algorithm was used to trace risk-evolution chains, followed by a comprehensive evaluation integrating safety and efficiency. Key findings include: (1) significant and rapid safety decay being accompanied by swift recovery, and the following vehicle exhibiting significant feature differences across recovery patterns; (2) safety decay in HV–AV being slower than that in HV–HV, with HV–AV demonstrating a close and conservative driving strategy during the risk-evolution process, with HV–AV meanwhile performing worse in safety and efficiency compared with HV–HV, highlighting the interaction discordance between HV and AV; (3) excessive deceleration of HV being the primary trigger for safety degradation in AV–HV, and AV demonstrating effective adaptation to safety decay, promoting slight safety reduction with rapid recovery, thereby balancing safety and efficiency. These findings reveal principles of risk evolution in mixed-driving environments, providing a novel analytical framework for mixed-driving safety.

Traditional expressway tunnel lighting systems usually lack sufficient consideration of the dynamic luminance demand arising from real-time traffic conditions and smooth luminance transitions between different tunnel lighting sections. These shortcomings may impede the effective accommodation of drivers’ visual adaptation during tunnel entry and transition and lead to unnecessary energy consumption due to over illumination. To address these issues, we propose in this study a deep learning based framework for tunnel luminance demand prediction. The standard luminance demand for each tunnel lighting section is first calculated according to the current national tunnel lighting standards. Then, a stepwise luminance demand division between tunnel lighting segments is achieved by using a combination of piecewise cubic Hermite interpolating polynomial and circuit segmentation, enabling a gradual transition of luminance between adjacent segments. Furthermore, taking the influence of variables such as environmental light intensity, traffic volume and vehicle speed on tunnel luminance demand into account, an attention-based dual LSTM model is developed to predict real-time tunnel luminance demand, so as to achieve intelligent “vehicle-following dimming” effect. The proposed model is validated by a case study from the Taolu Gou Tunnel on the Zhangcheng Expressway in Chengde, China. The model achieves an overall average R ² exceeding 0.86 across all lighting segments, with all error metrics maintained at low levels. Compared to the BPNN, RNN, and conventional LSTM models, the prediction accuracy is improved by 13.02%, 19.66%, and 11.77%, respectively, which demonstrates its superior prediction performance within the investigated tunnel scenario and indicates its potential applicability under similar operational and environmental conditions.

Resilience represents a crucial safety paradigm characterizing system risk evolution under specific perturbations, yet its application to microscopic driving safety needs further exploration. Analogous to resilience evolution patterns, generalized driving risk progression exhibits distinct safety decay and recovery phases, thereby establishing theoretical foundations for driving safety resilience analysis. This study develops a resilience-centric framework for quantifying dynamic car-following risk evolution. The car-following pairs were selected from highD natural driving trajectory dataset. Following risk quantification and clustering, 1,676 samples demonstrating complete “safe-dangerous-safe” transition process were identified. Through decomposition of car-following resilience into safety decay and recovery phases, we derived five interpretable resilience features and modeled their functional dependencies on driving variables via an integrated machine learning and SHapley Additive exPlanations (SHAP) analytical framework. Key findings reveal: 1) the CatBoost demonstrated superior fitting performance, achieving a mean absolute percentage error less than 10% across all resilience features; 2) the proposed Low-Rank Polynomial SHAP Fitting (LRP-SF) captured the nonlinear relationships between driving variables and resilience features, quantifying both directional influences and threshold effects; and 3) driver risk perception exhibits phase-dependent variability, with a stimulus-response mechanism governing safety evolution dynamics. The rapid safety deterioration and hazardous states act as triggers for recovery processes. This study further examined resilience threshold validity, car-following variable interactions, and inter-feature correlations. The potential application of LRP-SF in car-following safety control was also anticipated. This study offers methodological advancements for advanced driver assistance systems development and establishes a novel analytical paradigm for driving safety research.

Lane-changing is a typical scenario of driver interaction, which causes a multitude of traffic crashes as well. Therefore, clarifying the lane-changing interaction process is of great significance for driving safety. Although extensive research has been carried out in this respect over the past decades, there are still unsolved challenges ranging from data collection to model construction. Moreover, the influence of human factors on the interactive decision-making and behavior is not fully taken into account. In this study, we start with behavioral data collection based on a two-vehicle simulated driving experiment, followed by the construction of a recurrent multi-time step game modeling framework, in which driving styles of both drivers are considered. More specifically, we design 8 lane-changing experimental scenarios at urban expressways, and conduct a two-vehicle simulated driving experiment and combine it with a driving style questionnaire (MDSI) to collect the interactive driving behavior data and driving style information of 40 experimenters, based on which we analyze the influence of different factors on these drivers' lane-changing interactive behavior. Thereafter, we select the significant influencing factors to construct a payoff function and propose a dynamic game model that can explain the mechanism of the interactive decision from a microscopic perspective. Compared with previous lane-changing game models, the proposed model incorporates driver styles into the game payoff function and establishes a recurrent multi-time step interactive decision-making framework by a dynamic game tree to capture the interdependence of drivers' interactive decisions. The model is solved by parallel computing methods. Comparing the vehicle trajectory prediction results calculated by different models, we find that the error of the multi-time step model considering driver styles is 57.94% lower than that of the single-time step model without driver styles, which implies that the model proposed in this study can better simulate the process of drivers' actual lane-changing interaction behavior by considering the drivers' prediction planning cycle and driving styles. Especially, the driver styles have a more significant effect on reducing the model errors (42.87%), which however, was often ignored in the existing studies.