Shuguang Li

With the transformation of artificial intelligence from a "rule-driven" to a "data-driven" paradigm, the end-to-end architecture has become a frontier direction in the development of intelligent driving systems. However, the "black-box" nature of its decision-making poses significant challenges to the current admission management framework based on modular design, resulting in insufficient regulatory foundations and incomplete policy coverage. To address these risks, a scenario-based dynamic risk quantification model is developed to provide a scientific basis for accurate assessment and hierarchical supervision. A full-chain safety management mechanism covering four dimensions is further established to systematically control risks throughout the entire lifecycle of the technology. Finally, three core policy optimization measures are formulated to facilitate the transition from "static access" to "dynamic access + continuous safety assessment", to strengthen the supporting standards and technical systems, to define the mechanisms for hierarchical accountability and social risk sharing, and to establish a governance framework for the safe and orderly deployment of end-to-end intelligent driving systems.

Multimodal image registration techniques, critical for advancing the measurement science of infrared and visible images, focus on aligning the same feature points across diverse modalities to achieve comprehensive and accurate instrumentation. However, it is still challenging to achieve high accuracy and robustness for the infrared and visible image registration with complex deformations and significant modal differences. In this study, we propose a consistency-constrained bidirectional registration network, which enhances the stability of the registration process by capturing both forward and backward registration details. Furthermore, we develop a synergistic unimodal and multimodal supervision framework, which is designed for learning and imposing cross- and intra-modal registration constraints. Experiments on several benchmark datasets have demonstrated that the proposed method not only improves registration accuracy, but also exhibits excellent stability. Specifically, our method achieves 1.382, 2.680, and 2.557 endpoint errors on the RoadScene, FLIR, and MSRS datasets, respectively, resulting in error reductions of 13.24%, 5.20%, and 11.61% over existing methods. The proposed method surpasses existing techniques in handling image registration with complex deformations and significant modal differences, as it effectively extracts the commonality features between the infrared and visible images, thereby improving the registration accuracy.

Semantic Change Detection (SCD) based on remote sensing images can provide an effective solution for large-scale land use monitoring. Existing change detection networks face limitations due to their limited receptive fields, which make it difficult to provide a comprehensive and consistent response to changes in similar large objects. Moreover, these limitations often cause the network to ignore weak changes localized in the images. To address these limitations, we propose CG-SPNet, a decoupled multi-task architecture that integrates three key components, including the multiscale positional similarity pyramid (MSSP), the symmetric fusion change detection unit (SFCD), and the change-guided feature interaction module (CGM). MSSP module captures positional correlation of semantic features at multiple scales by introducing an attention pyramid during pooling down-sampling. SFCD utilizes a convolutional attention mechanism to enhance local features and highlight weak changes in dual images. And the symmetric fusion approach reduces the network’s sensitivity to timing. CGM is used to address the challenge of effective feature interaction, which combines cosine similarity loss to leverages cross-attention to compute the positional relevance of change features to embed a priori information for semantic features. Through rigorous experiments and analyses, we have successfully validated the essential components of CG-SPNet. It achieves state-of-the-art performance on the SECOND dataset as well as our self-built CZWZ dataset.

Traffic accidents due to fatigue account for a large proportion of road fatalities. Based on simulated driving experiments with drivers recruited from college students, this paper investigates the use of heart rate variability (HRV) features to detect driver fatigue while considering sex differences. Sex-independent and sex-specific differences in HRV features between alert and fatigued states derived from 2 min electrocardiogram (ECG) signals were determined. Then, decision trees were used for driver fatigue detection using the HRV features of either all subjects or those of only males or females. Nineteen, eighteen, and thirteen HRV features were significantly different (Mann–Whitney U test, p < 0.01) between the two mental states for all subjects, males, and females, respectively. The fatigue detection models for all subjects, males, and females achieved classification accuracies of 86.3%, 94.8%, and 92.0%, respectively. In conclusion, sex differences in HRV features between drivers’ mental states were found according to both the statistical analysis and classification results. By considering sex differences, precise HRV feature-based driver fatigue detection systems can be developed. Moreover, in contrast to conventional methods using HRV features from 5 min ECG signals, our method uses HRV features from 2 min ECG signals, thus enabling more rapid driver fatigue detection.

Depth estimation is an important part of the perception system in autonomous driving. Current studies often reconstruct dense depth maps from RGB images and sparse depth maps obtained from other sensors. However, existing methods often pay insufficient attention to latent semantic information. Considering the highly structured characteristics of driving scenes, we propose a dual-branch network to predict dense depth maps by fusing radar and RGB images. The driving scene is divided into three parts in the proposed architecture, each predicting a depth map, which is finally merged into one by implementing the fusion strategy in order to make full use of the potential semantic information in the driving scene. In addition, a variant L1 loss function is applied in the training phase, directing the network to focus more on those areas of interest when driving. Our proposed method is evaluated on the nuScenes dataset. Experiments demonstrate its effectiveness in comparison with previous state of the art methods.

The development of a machine-learning method with high accuracy, high generalization, and strong robustness for evaluating battery health states is essential in the field of battery health management. In this work, the data-driven stacking regressor (SR) method with a two-layer diagnostic framework was proposed to estimate the state of health (SOH) and predict the remaining useful life (RUL). Five individual estimators were merged in the first layer, including bagging, gradient boosting regression (GBR), support vector regression (SVR), Hist-GBR, and AdaBoost, and linear regression (LR) was used in the second layer to construct the SR model. The SR model produces highly accurate results without the requirement of excessive parameter adjustment. Fifteen batteries from the NASA dataset were used for our experiments, resulting in rather low values of average root mean square error (ARMSE) and relative error (RE) for the SOH estimation and RUL predictions of the different batteries, demonstrating the superiority of the SR model.

Vision-based localization techniques and detection technologies are key algorithms for the localization and navigation of unmanned vehicles. Especially in scenarios where GPS signals are missing, Simultaneous Localization and Mapping (SLAM) techniques that rely on vision, inertial navigation system (INS) and other sensors have important applications. Among them, vision combined with the IMU SLAM system has the advantage of realistic scale, which is lacking in monocular vision and computational power compared to multi-visual vision, so it is suitable for application in an unmanned vehicle system. In this paper, we propose a fusion localization algorithm that combines a visual-inertial SLAM system and map road information, processing road information in a map under structured roads, and detecting lane lines and locating its local position by a monocular camera, applying a strategy of position prediction and update for map-SLAM fusion localization. It solves the problem of accumulating errors in a pure SLAM system without loopback and provides accurate global-local positioning results for unmanned vehicle positioning and navigation.

Electromechanical transient (EMT) and transient stability hybrid simulations can make full use of their fast-solving ability and high-precision computing ability to accurately simulate the dynamic characteristics of large-scale AC/DC systems. This paper selects the bus as the hybrid simulation interface bus in the AC system near the DC system, which can effectively improve the simulation accuracy of the AC/DC system interaction. At the interface bus, Norton equivalent and Thevenin equivalent are applied to the EMT simulation side and electromechanical transient simulation side, respectively. The three-phase simulation data and three-sequence component simulation data are extracted respectively, and the serial iteration method is used for hybrid simulation calculation. This paper takes Zhangbei multi terminal AC/DC system as an example, compares the hybrid simulation results with the EMT simulation results, and tries to verify the accuracy and effectiveness of the hybrid simulation method.