Scenario-based testing is an important part in intelligent vehicle (IV) development. The data volume of collected test scenarios is extremely large, and directly using all collected scenarios to test IVs will lead to extremely low testing efficiency. To solve this problem, a criticality assessment model (CAM) for IV test scenario based on interactive field feature (IFF) and hypergraph learning is proposed to quantify the test scenario criticality to improve the test efficiency. The IFF is constructed based on the potential field-based method to integrally consider the multidimensional coupling of scenario elements. In addition, the interaction between the fields generated by the ego vehicle and the driving environment is modeled based on Delaunay triangulation discretization method to accurately quantify the driving environment risk to the ego vehicle. The node and hyperedge of the hypergraph are used to model the individual dynamic evolution and group interaction characteristics of vehicles, respectively. Subsequently, a hypergraph learning network is constructed to extract features from the built hypergraph, IFF and traffic elements. Finally, the effectiveness, reasonableness and accuracy validation experiments are designed to validate the proposed CAM. Ablation experiment results show that the proposed IFF and hypergraph learning network enhance the CAM accuracy. The reasonableness validation results show that the proposed CAM can better find critical test scenarios than time-to-collision and time-head-way methods. The accuracy of the proposed CAM is compared through four real validation scenarios in the proving ground. The comparison results show that the proposed CAM can accurately output the quantified test scenario criticality.

Trajectory planning is crucial for ensuring the safety of intelligent vehicles (IVs) in autonomous driving, especially in complex traffic environments where perception and control uncertainties increase collision risks. Existing risk-aware planning methods often fail to capture the full scope of uncertainties through risk assessment and struggle to consistently integrate risk information throughout the trajectory generation process. Risk factors are often overshadowed by smoothness and dynamic feasibility considerations, leading to excessively high-risk trajectories that fail to meet human driving expectations for caution. To address these challenges, we propose a cognitive risk-aware hierarchical trajectory planning method that adaptively regulates driving caution based on cognitive risk. We introduce cognitive risk assessment using two fields: the anisotropic objective risk field, which accounts for the size and comprehensive motion uncertainty of surrounding obstacles, and the driving cognition field, which considers the IV motion trends and reflects its proactive cognition of objective risks. By fusing these fields, we calculate cognitive risk and incorporate it into a planning framework that combines trajectory search and optimization. By explicitly considering cognitive risk constraints in both stages, the method achieves complete cognitive risk awareness, generating adaptive, safe, and dynamically feasible smooth trajectories. Experimental results in various scenarios demonstrate the effectiveness and superiority of the proposed method. By incorporating cognitive risk assessment, the IV exhibits more cautious driving behavior, aligning with human expectations. Compared to three state-of-the-art methods, our method improves the minimum time-to-collision by over 20%, reduces lane-crossing time by more than 10%, and decreases the average yaw rate by over 12%. In summary, the proposed method ensures higher safety, improves lane-changing efficiency and smoothness.

Longitudinal control systems in intelligent connected vehicle (ICV) significantly reduce driving workload; however, most existing designs overlook the driver’s psychological perception of car-following safety. This neglect can elevate physiological stress and erode trust in the system. To address this, this paper proposes a human-centered Connected Adaptive Cruise Control (C-ACC) strategy that integrates the driver’s individualized cognitive safety analysis into the control loop within a connected environment. Using naturalistic driving data that jointly capture traffic context, vehicle states, and driver physiology, we construct an individualized safety boundary in the speed–distance domain. The boundary is parameterized by a Sigmoid function fitted to physiological activation features and is incorporated as a soft constraint within a Model Predictive Control (MPC) framework. Crucially, the controller leverages Vehicle-to-Vehicle (V2V) communication to optimize tracking performance while proactively limiting states associated with elevated activation. The strategy was validated through driver-in-the-loop experiments using two real vehicles enabled with V2V communication. Experimental results demonstrate that, compared with baseline strategies, the proposed approach effectively reduces physiological activation and psychological tension while ensuring ride comfort and control stability, thereby enhancing the driver’s perceived safety and trust. This framework provides a principled path for personalized C-ACC design based on cognitive state estimation and is scalable to broader connected automated driving applications.

Jiayi Han, Liang Du, Hongwei Du, Xiangguo Zhou, Yiwen Wu, Yuanfang Zhang, Weibo Zheng, Donghong Han. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Registration of intraoperative ultrasound (iUS) with preoperative CT represents a significant yet challenging task in the context of radiation-free spinal surgical navigation. The presence of thickness response artifacts in US images poses a considerable obstacle to the accurate extraction of bone boundaries. Furthermore, US-CT registration typically necessitates the detection and correspondence of high-quality landmarks at the initial stage. This can be accomplished by surgeons who have undergone extensive training in the localization of standard spinal US views, enabling them to identify key vertebral landmarks for subsequent precise registration. In this paper, we propose a fully automated iUS registration method that employs a limited number of spinal US views as observation objects. Specifically, three-dimensional vertebral meshes segmented from the preoperative CT images are superimposed on the US images and then fed to the reinforcement learning (RL) agent for sequential decision-making. The proposed method achieves fully automatic US-CT registration without relying on prespecified initialization. This is achieved by training the agent to approach bone surfaces on several randomly selected 2D US views. The instability of RL-based iUS registration is primarily attributable to the difficulty of correlating long-range information within the neural network. To address this issue, we propose a Field of View Separation (FoVS) module. The proposed approach employs separate encoders for US and mesh images, followed by cross-attention aggregation, which facilitates information flow between non-adjacent pixels. This approach enables pretraining of feature extraction on distinct encoders and the application of supplementary loss for enhanced feature matching precision, thereby significantly improving the learning capability and stability of the network. Furthermore, a refinement module is introduced to correct the results of the RL registration, which improves the stability of the registration process. To ascertain the efficacy of each module, action, and auxiliary task, comprehensive experiments are conducted. The results demonstrate that the performance of the RL agent is enhanced by the associated modules and auxiliary tasks. The registration exhibited an angular error of$8.83 \; \pm \; 4.69$degrees and a translational error of$3.34 \; \pm \; 1.42$mm, achieving the state-of-the-art (SOTA) results. It is noteworthy that fine-tuning the model prior to the surgical phase can significantly reduce the registration error, which is a promising outcome for its clinical translation.Note to Practitioners—The objective of this study is to address the issue of image registration using iUS in conjunction with preoperative CT scans in the context of spine surgery. The current 2D/3D image registration methods are constrained by several limitations. Firstly, they often exhibit reduced accuracy, and require high-quality images in substantial quantities. Secondly, there is a lack of effective mechanisms to rectify errors identified after the registration process. This paper proposes a fully automated registration framework based on RL, which incorporates image rendering and mesh clipping to enable continuous adjustment of the pose of 3D data, thereby facilitating 2D/3D registration. The framework employs the distinctive attributes of iUS images and incorporates a refinement module to evaluate registration accuracy, thereby facilitating the rectification of any registration issues. The proposed framework was tested on both sheep lumbar subjects and human lumbar phantoms, demonstrating the highest level of performance to date and indicating its potential for integration into surgical navigation systems.

The ultrasonic radar is a crucial part of the automatic parking system (APS) and requires thorough testing to ensure safety. The field operational test (FOT) is expensive and inefficient, making simulation testing the preferred method for evaluating autonomous driving systems. However, a gap remains between the performance of current simulation ultrasonic radar models and the real-world ultrasonic radars in real driving environments. To address these challenges, an ultrasonic radar in-the-loop accelerated test method for APS is proposed. First, a high-fidelity ultrasonic radar in-the-loop test bench is established and validated. To reduce the high cost and time duration of hardware-in-the-loop (HIL) testing, a Bayesian optimization (BO)-based accelerated testing framework is designed, minimizing test numbers by using the prior information to guide the search process. Finally, experiments are conducted using the ultrasonic radar in-the-loop test bench to compare the Bayesian accelerated test with the combination test (CT) for the automatic parking algorithm. The test results indicate that the ultrasonic radar in-the-loop test bench demonstrates greater authenticity and accuracy than the simulation software, while the HIL testing offers improved efficiency and safety compared to the FOT. Moreover, the efficiency of the Bayesian accelerated test is 4.9 times greater than that of the CT while still ensuring adequate scenario coverage.

To ensure that the camera functions (CFs) of intelligent vehicles can operate normally in rain and fog, sufficient tests for CFs are indispensable. In this work, the synthetic image generation model (SIGM) is proposed to generate images to synthesize rain and fog effects for testing CFs. First, to synthesize the rain blur effects on raw images, the geometric-physical fusion model is proposed based on the properties of raindrop microparticles and vehicle features. In this model, the position distribution of raindrops is determined by the geometric model part, and the image pixel effects are realized by the physical model part. Second, the atmospheric scattering model is used to synthesize the fog effect and contains two parts: an airlight estimation model to estimate airlight values and a transmittance estimation model, where the corrected image depth and fog microphysical information based on the Mie theory are used to determine the fog transmittance. Finally, SIGM is validated through real rain and fog in image quality, testing CFs with synthetic images and videos. The result absolute error sum obtained with SIGM (0.45) is smaller than that of the simulation software (0.57) and another synthetic method (0.65), and all unrecognized targets identified by SIGM match the real test results, whereas both the simulation software and another synthetic method fail to accurately test CF. The experimental results indicate that the proposed SIGM has higher authenticity and reliability than testing using the simulation platform and has higher efficiency, controllability, repeatability, and safety than testing using real vehicle.

Driving risk prediction is important for the development of intelligent vehicles (IVs), and the rise of human-like driving requires a driving risk prediction system to match the subjective risk cognitive characteristics of human drivers. In this study, a subjective driving risk prediction model (SDRPM) for IVs is proposed and applied to lane-changing (LC) conditions. We regard the cognitive risk of the human driver as the coupling result of environmental objective risk and driver subjective cognition at the spatial and temporal scales. Then, an objective anisotropic risk field was built to describe risks in traffic environments, and a subjective spatiotemporal cognition field to describe the cognitive characteristics of human drivers. By combining the two fields, the spatiotemporal distribution features of the human driver’s cognitive risk were obtained and used as model inputs, and SDRPM output the predicted subjective driving risk level (SDRL) of the human driver. To quantify the SDRLs, six participants were recruited to watch 1,213 driving videos and report the SDRLs they cognized. Participants were provided with 360-degree driving video around the test vehicle and virtual reality glasses to ensure the reliability of the obtained SDRLs. Verification results showed that SDRPM has good predicted accuracy and long advance predicted time, with 97.53% predicted accuracy, which can reach 95.06% at 2 s before the LC point. Compared with six state-of-the-art models, SDRPM can improve the predicted accuracy while reducing the dimensions of input features. In summary, SDRPM can ensure the driving safety, and improve user acceptance and trust in IVs.

Facial expression recognition (FER) remains a challenging task despite years of effort because of the variations in view angles and human poses and the exclusion of expression-relevant facial parts. In this work, we propose to learn effective Global receptive field and Class-sensitive metrics for FER, namely GCNet which contains a Class-sensitive metric learning module (CSMLM) and mobile dilation modules (MDMs). CSMLM fully takes advantage of the variation in human faces to extract class-sensitive and spatially consistent features to improve the effectiveness of FER. MDM utilizes cascaded dilation convolution layers to achieve a global receptive field. However, directly adding a dilation convolution layer to a given sequence of convolution layers may face the gridding problem, which leads to sparse feature maps. In this work, we find the upper bound of the dilation rate of the additional convolution layer that avoids the gridding problem. Experiments show that the proposed approach reaches state-of-the-art (SOTA) performance on the RAF-DB, FER-Plus, and SFEW2.0 datasets.

The intelligent connected vehicle (ICV) is an organic combination of connectivity and intellectuality, making it possible to exchange information and drive autonomously. How to assess and describe the traffic risk for the ICV is a fundamental and crucial problem. Unlike the existing studies that only obtain instantaneous and quasi-static risk fields due to the separation between space and time, this study proposes a novel Spatial-Temporal Risk Field (STRF) that represents the dynamic driving risk from the perspective of time-space coupling for ICVs in dynamic traffic is proposed. In order to generate the field for the different traffic elements, the influential traffic elements are divided into concrete elements and abstract elements according to their features and influencing mechanisms. And then, A biased sweeping method (BSM) is developed for concrete elements, and a modeling method based on Gaussian distribution is developed for abstract elements. The results of the STRF for a traffic scenario are visualized and analyzed in two forms: complete form and slice form. The results show that the STRF can precisely describe the risk distribution of each traffic element in a specific and characteristic shape according to the spatial-temporal situation. Additionally, this study also provides an application example of the STRF in trajectory planning to demonstrate the applicability and availability of the STRF. The STRF-based trajectory planning method greatly benefits from the STRF and shows the potential ability of flexible and personalized trajectory planning. The STRF proposed in this paper can express the dynamics and continuity of the spatial-temporal risk and describe the instantaneous risk by extracting the spatial risk distribution at a certain moment. The STRF can be used in risk assessment, decision making, trajectory planning, driving behavior modeling, and automatic testing.

With the development of intelligent vehicles, more research has focused on achieving human-like driving. As an important component of intelligent vehicle control, car-following control should ensure safety, tracking, comfort while considering the acceptance of human drivers. In this paper, we propose a car-following control strategy$\boldsymbol {\pi }_{ \boldsymbol {Hybrid}}$based on a hybrid of reinforcement learning (RL) and supervised learning (SL). RL is used to achieve multi-objective collaborative optimization in car-following control, and SL is used to achieve human like car-following. Through the complementary advantages of the two learning methods,$\boldsymbol {\pi }_{Hybrid}$can achieve high performance car-following while matching the personalized car-following characteristics of human drivers. RL is used as the main framework of$\boldsymbol {\pi }_{Hybrid}$. In addition, the personalized car-following reference model (PCRM) of human drivers based on Gaussian mixture regression, and the motion uncertainty model of preceding vehicle (MUMPV) based on the sequence-to-sequence network are established and incorporated into the RL framework. PCRM can lead$\boldsymbol {\pi }_{Hybrid}$to learn the different characteristics of human drivers, and improve the anthropomorphism of$\boldsymbol {\pi }_{Hybrid}$; MUMPV enables$\boldsymbol {\pi }_{Hybrid}$to consider the dynamic changes of the traffic environment and to become more robust.$\boldsymbol {\pi }_{Hybrid}$is trained and tested on High D dataset, and the generalizability verification is based on the self-built real vehicle data collection platform. The results show that$\boldsymbol {\pi }_{Hybrid}$can match human drivers’ personalized car-following characteristics and can outperform human drivers in safety, comfort, and tracking of the preceding vehicle.

With the development of intelligent vehicle technology, many studies have been focused on developing human-like trajectory planning methods for automated driving systems. Although data-driven methods are widely used for human driver behavior learning, there have been fewer studies on realizing human-like trajectory planning by using the generation mechanism of driving behavior, especially under curve conditions, where the lane centerline has been denoted as a reference trajectory. In this paper, thirty-two skilled drivers were recruited to collect data under different curve conditions on a self-designed driver-in-the-loop system. The collected data are processed by dynamic time warping, trajectories with different lengths are warped and the abnormal data are removed. Based on the warped data, common characteristics and differences between left and right turning trajectories are compared and explored from the perspectives of drivers’ demand for turning performance and their visual attention mechanism. Then, by introducing the driver preview mechanism, two features with a strong ability to represent the generation mechanism of the driver’s curve driving behavior are introduced. Finally, the preview-based human-like trajectory planning model (PHTPM) is proposed, and it is verified and analyzed by comparative tests and generalizability tests. The results show that the introduction of the driver preview mechanism enables PHTPM to match the characteristics of skilled drivers accurately on left turnings and outperform them on right turnings.

Intelligent vehicle testing has been a hotspot in the field of intelligent vehicles. Due to the multi-dimensional parameters and the continuity of testing scenarios, a complete test of all scenarios requires a large amount of manpower and several material resources. In critical scenarios, deficiencies in an intelligent vehicle’s performance and defects of an algorithm can be exposed. Therefore, increasing the search efficiency and coverage of critical scenarios is key in improving scenario-based intelligent vehicle testing technology. In this study, a critical scenario search method for intelligent vehicle testing based on the social cognitive optimization (SCO) algorithm is proposed. This method has two main parts: global search and local search. The global search, based on the modified SCO algorithm, integrates the density peak clustering (DPC) algorithm and the cooling scheduling function, and aims to find all local aggregation areas of critical scenarios in a logical scenario space. The local search applies a multi-dimensional convolution algorithm to the global search results to find critical scenarios near the local aggregation areas. Finally, a longitudinal automatic driving algorithm is tested using the proposed method under a specified logical scenario in a simulation environment. The test results show that the proposed method can improve both the search efficiency and coverage of critical scenarios.

Although considerable progress has been achieved regarding the transformers in recent years, the large number of parameters, quadratic computational complexity, and memory cost conditioned on long sequences make the transformers hard to train and implement, especially in edge computing configurations. In this case, a dizzying number of works have sought to make improvements around computational and memory efficiency upon the original transformer architecture. Nevertheless, many of them restrict the context in the attention to seek a trade-off between cost and performance with prior knowledge of orderly stored data. It is imperative to dig deep into an efficient feature extractor for point clouds due to their irregularity and a large number of points. In this paper, we propose a novel skeleton decomposition-based self-attention (SD-SA) which has no sequence length limit and exhibits favorable scalability in long-sequence models. Due to the numerical low-rank nature of self-attention, we approximate it by the skeleton decomposition method while maintaining its effectiveness. At this point, we have shown that the proposed method works for the proposed approach on point cloud classification, segmentation, and detection tasks on the ModelNet40, ShapeNet, and KITTI datasets, respectively. Our approach significantly improves the efficiency of the point cloud transformer and exceeds other efficient transformers on point cloud tasks in terms of the speed at comparable performance.

Human-machine shared control has become an effective cutting-edge approach to enhance driving safety and assist in the transition from manual to autonomous driving. However, driving authority allocation when conflict occurs between a human driver and a machine still represents an intractable problem. In order to solve this problem, this paper proposes an adaptive steering torque coupling framework to achieve the human-machine shared control with conflict resolution. Additionally, a driver trajectory prediction method considering human-machine interaction is established based on the extended state observer and long short-term memory network. In addition, the human-machine shared steering system is modeled based on a non-cooperative dynamic game by the prediction model method for path-tracking, and the obtained solution is given according to the Nash equilibrium. Particularly, a dynamic load allocation approach is designed to resolve human-machine conflict and relax the driver. In order to verify the strategy formed based on the proposed framework, a driver-in-the-loop experiment is conducted. The experimental results reveal that the proposed strategy can effectively reduce the human-machine conflict torque and ensure a driver has absolute control authority. Furthermore, the proposed strategy can also transfer the driving workload between the driver and machine, provide driving experience, or improve driving comfort when there is no conflict.

There have been explosive developments in automatic driving in recent years. Several kinds of self-driving vehicles have been introduced, but drivers are generally required to stay alert and put their hands on the steering wheel during automatic driving, which means the vehicle is in a state of human-machine cooperative control. Hence personalization is needed to reduce conflicts between driver and machine. To achieve personalized automatic driving, a combined hierarchy learning framework (CHLF) based on gated recurrent units (GRU) and safety field is developed in this paper. Learning-based methods are gaining attention in the development of automatic vehicles. However, automatic driving based on neural networks is risky due to poor interpretability. To overcome this limitation, we divided the network into three layers according to function to achieve a hierarchy. We combine data- and mechanism-oriented methods to make the CHLF reliable and stable. Lane-changing (LC) driving data collected from real-vehicle field experiments are used to train the CHLF, which learns from the data to capture human driver behavior in LC. To verify the performance of the CHLF, a group of driver-in-the-loop experiments is conducted based on a driving simulation test bench. The results show that the CHLF can reduce driver steering output torque compared to the conventional method, which means there is less conflict between driver and machine during automatic LC, and the driver is more relaxed.