The LCL motor inverter is modeled using the complex vector in the synchronous rotating reference frame (SRF), and the influence of the overall phase margin and coupling performance of the system under different control strategies is thoroughly investigated. The analysis reveals that the inverter has a coupling relationship between the dq axes, and the active damping feedback branch also intensifies the coupling phenomenon. To address these issues, a complex vector decoupling strategy is proposed. The decoupling components for the active damping feedback branch and the current loop system are sequentially designed. The decoupling elements are applied to both the feedback branch and the output current of the PI controller respectively, which not only effectively eliminate the dynamic coupling but also enhances the system's dynamic performance. Finally, the effectiveness of the control strategy is verified by comparing the simulation and experimental results with the traditional decoupling control.

Conducting a thorough driving risk assessment is important for the driving safety of autonomous vehicles. In this paper, the existing driving risk assessment methods are divided into three categories, namely, the single objectoriented methods, the reachability setbased methods, and the potential fieldbased methods. In order to conduct a comprehensive comparison of these methods and reveal their distinct characteristics and applicability, the paper proposes five evaluation dimensions, including realtime capability, the duration of the valid prediction horizon, application feasibility, the inclusion of various risk sources and adaptability in different scenarios. The research gaps and potential future research directions in driving risk assessment for autonomous vehicles are analyzed and prospected.

With the advancement of autonomous driving technology, the accuracy and reliability of object detection have become increasingly crucial. Deep learning, as a core component of autonomous driving systems, significantly influences the safety and stability of these systems by estimating the uncertainty in predictive results. The paper summarizes the application of deep learning uncertainty estimation in autonomous driving object detection and discusses the significance of an effective uncertainty evaluation system. Firstly, the paper introduces the fundamental theories of deep learning uncertainty estimation, including Bayesian neural networks, Monte Carlo methods, and ensemble learning. These methods quantify model prediction uncertainty in different ways, providing autonomous driving systems with richer information. Secondly, the paper delves into the application of uncertainty estimation in autonomous driving object detection. Through case studies, it demonstrates how uncertainty information can be used to improve detection accuracy, especially in complex environments and extreme conditions. In these scenarios, uncertainty estimation provides decision support, helping the system avoid potential risks. Lastly, the paper summarizes the evaluation metrics for uncertainty estimation in autonomous driving object detection, considering both the model's predictive performance and the accuracy of the uncertainty estimation.

In the visual perception task of autonomous driving, it is crucial to accurately and quickly extract the cognitive and accidental uncertainties to effectively resolve the Safety of the Intended Functionality (SOTIF) issues associated with autonomous driving. In traditional methods such as Monte Carlo dropout and deep ensembles, uncertainty is estimated by sampling the prediction results of different submodels, which slows down the estimation and tends to occupy a large amount of memory in the processor during the model inference stage. A fast Monte Carlo dropout method and a technique for correcting subsequent detection results are proposed to address the issues of slow estimation of uncertainty in Monte Carlo dropout and the selection of subsequent detection results. This method uses a multihead mechanism to replace the traditional multiple sampling mechanism in Monte Carlo dropout, thereby saving time in both sampling and inference throughout the uncertainty estimation process.

Aiming at acceptance criteria for automated driving, this paper reviews the best practices in relevant regulations, standards, and evaluation methods, identifying five safety concepts and their interrelationships. The paper focuses on the concept and research status of behavioral safety, centered around "reasonably foreseeable and preventable” behaviors. By combining scenario data statistics with the driver's emergency response mechanism, a quantitative research framework is proposed for reasonably foreseeable and preventable situations. Finally, combining traffic accident data and practical experience, the paper provides a method for using behavioral safety in autonomous driving evaluation, along with a closedloop certification and approval process based on this concept. The research in this paper serves as a reference for authorities, third parties, and R&D companies to establish relevant R&D, testing, and processes centered around behavioral safety.

To avoid unnecessary interventions by the driver assistance system, this paper combines collision risk and driving maneuverability to introduce the concept of a risk assessment zone in longitudinal following scenarios. The boundary of this zone is determined based on the normal distribution characteristics of the driving data. Subsequently, a new humanmachine codriving longitudinal driving rights allocation strategy is proposed, which takes the inverse time to collision (TTCi) as the basis for judgment. If the TTCi exceeds the threshold value, the upper boundary of the risk assessment zone represents the maximum deviation in driving maneuverability. The control rights of the assistance system are allocated according to the deviation in the driver's maneuverability. By combining Prescan, Matlab/Simulink and the Logitech G29 driving simulator, a driverintheloop simulation platform was constructed. The platform simulated the reduced driver maneuverability due to distracted driving, thereby verifying the effectiveness of the strategy. The results show that the proposed humanmachine codriving strategy can effectively prevent collisions caused by reduced driver maneuverability under highspeed road following conditions.

Aiming at the lack of effective methods for testing abnormal signals during the operation of unmanned vehicles, the paper focuses on signal anomalies caused by environmental disturbances in reliability driving tests. By using the correlation of signals from multiple sensors in both the time domain and spatial domain, a crossmathematical model is established based on the multisensor data. The signals collected from sensors are assigned as the row elements and the sensors as the column elements within the signal matrix. This numerical method transforms the original multisensor signals into a parameterized signal matrix model. A method combining matrix completion and deep matrix decomposition fusion (MC+DMF) is proposed to recover certain abnormal signals resulting from environmental disturbances. According to the forward propagation characteristics of the neural network, dimensionality reduction is applied to the row vectors (data collected by individual sensors at time i ) and column vectors (sensor arrays) in the original matrix. This process reduces the computational load during feature extraction from the distorted signal matrix. Additionally, the Hadamard product is used to regularize the MC+DMF loss function after feature extraction to avoid overfitting. The proposed method is applied on the SODA10M and KITTI public datasets, and comparing with traditional approaches, such as the single matrix factorization (MF), probability matrix factorization (PMF) and BiasSVD, the experiments using root mean square error (RMSE) show that the method can effectively detect abnormal sensor signals caused by vibration interference during driving. The results show that the MC+DMF method can greatly reduce the data recovery error and time. Compared with the probability matrix decomposition method, it achieves a 1% lower error rate and approximately 20.65% less recovery time.

Lane line detection is a key technology in the field of autonomous driving, and it currently faces many challenges. The sparsity of the lane line supervision signal, as well as factors such as occlusion and shadows in complex scenes, can affect detection accuracy and realtime performance. Based on this, this paper proposes a lane line detection model that integrates the CBAM attention mechanism and a line anchor feature aggregation module. The proposed algorithm achieves an accuracy of 96.19% and a comprehensive F1 score of 76.24% on the Tusimple and CULane datasets, respectively. Real vehicle tests show that the algorithm detects a frame rate of 67 fps, allowing for realtime detection in complex traffic scenarios and more effectively addressing the problem of lane line occlusion.

In existing visionbased intelligent wiper systems, the raindrop target detection model has a large number of parameters and excessive computational complexity, making it challenging to deploy in vehicle embedded devices. To address these issues, the paper proposes a lightweight raindrop target detection model, YOLOV5RGA. By integrating the RepVGG and GhostBottleneck modules to replace the convolution and C3 modules of the backbone network, we enhance the network's feature extraction capabilities while significantly reducing the parameters and computational load. Furthermore, adopting the Adam optimizer results in faster convergence and improves the average accuracy of the network model. Through experimental validation, compared with the YOLOv5s model, the YOLOv5RGA model achieves a 0.8% increase in average accuracy. Additionally, the number of model parameters is reduced by 48.5%, computation demand decreases by 35.2%, and the model size shrinks by 44.4%. The adoption of the lightweight raindrop target detection model effectively reduces hardware overhead and also facilitates model deployment.

In order to accelerate the gathering, sharing, development and utilization of ICV data resources, this paper proposes the establishment of a “public service platform for data interaction and comprehensive application of ICV” with a multicenter architecture comprising national, regional and enterprise levels. The platform is based on the standardized industry data collection and transmission protocols and extensively utilizes modern information technologies such as big data, cloud computing and blockchain. It can efficiently achieve the realtime collection, analysis and processing of data from tens of millions of vehicles, providing a basic support platform for promoting the mining and utilization of industry data. Based on the data resources collected and stored by the platform, we have explored comprehensive applications across multiple scenarios, including vehicle test and evaluation, safe operation monitoring, and data analysis and mining. These applications verified the platform's innovativeness and feasibility in actual use.