Chang Wang

Driver’s attention area recognition (DAAR) is crucial for autonomous vehicle takeover and detecting distracted driving. Currently, deep learning-based models for DAAR have achieved promising results. However, due to substantial variations in driving environments and individual driver characteristics, end-to-end deep learning models struggle to maintain stable generalization performance across different driving scenarios. To address these challenges, this paper proposes a driver attention area recognition framework based on improved Transformer and Kalman Filtering (DAARF-TK). DAARF-TK comprises two key components: a facial landmarks extraction module and a Transformer modeling module. The face landmarks extraction module includes an Adaptive Unscented Kalman Face Landmark Filtering (AUKFLF) algorithm, calibrating face landmarks in scenarios with large face occlusions, significant head rotations, and blurriness. An efficient attention mechanism and a Tokens Dimension Reduction Module are embedded into the Transformer to enhance the adaptability of the Transformer architecture to the real-time requirements of the DAAR task. Experimental results reveal that when trained on the Recorded dataset, DAARF-TK attains mAcc values of 94.13%, 91.67%, and 75.38% on the Recorded, 100-Driver, and DGW test sets, respectively. In addition, a real-time validation conducted on a driving simulator indicates that DAARF-TK operates at 37.68 frames per second (FPS), confirming its real-time capability in DAAR. These findings underscore the superiority of DAARF-TK and its potential to offer technical support for autonomous vehicle takeover.

Predicting driver gaze area is critical for intelligent driving systems to assess driver state for takeover decisions and distraction warnings. However, the stochastic nature of driver visual behavior has posed significant challenges for developing reliable predictive solutions. This study introduces a framework for predicting driver gaze areas during in-vehicle information system (IVIS) secondary tasks, leveraging a hybrid deep learning model with spatio-temporally distributed multidimensional features to classify whether drivers are looking at the touchscreen area. The framework comprises two key components: multi-source feature extraction and hybrid model construction. For feature extraction, we employ the i-Tracker, TokenHPE, and MediaPipe models to capture driver gaze, head posture, and hand motion dynamics, respectively. In model construction, the framework integrates the local receptive field advantages of convolutional neural networks (CNNs) and the long-term dependency modeling of bidirectional long short-term memory networks (BiLSTMs). Specifically, BiLSTM layers extract fine-grained temporal features from multi-layer CNN outputs, while an attention mechanism optimizes weight allocation to prioritize critical spatio-temporal patterns. This results in a CNN-BiLSTM-Attention hybrid model for driver gaze prediction. Experimental results demonstrate that the proposed model outperforms baseline architectures in gaze area prediction. Under a sliding window step (s) of 2 s and prediction interval ( $\Delta ~{t}$ ) of 1 s, it achieves an accuracy of 89.2%, showcasing robust performance in capturing dynamic visual behaviors. This study provides a novel technical approach for intelligent driving systems to proactively monitor driver attention and supports the development of distraction warning and human-machine collaborative control strategies.

The reflections of multi-colored lights involved on rainy nights present strong uncertainties and abruptness, resulting in a high rate of false and missed detections in existing methods. To solve this issue, this paper proposes a lane detection method based on memory and discretization mechanisms. Firstly, a Memory Fruit-fly-optimizer with Individual Differences (MFID) is innovatively proposed to drive Multi-threshold Otsu (MOtsu)-based multi-class segmentation of lanes, which is a high-dimensional optimization task with real-time and local optimal challenges, for capturing lane clues obscured in multi-intensity reflections, consequently reducing missed detections. Specifically, to solve the challenges inherent in the task, the MFID incorporates a novel memory mechanism to establish fast-converging initial conditions for real-time detection, while creatively considering individual differences to motivate multi-swarm optimization that mitigates local optima risks. After integration, the MFID-MOtsu is constructed for lane segmentation. Subsequently, a dynamic discretization mechanism is proposed to efficiently separate lane edges from interference edges, mitigating accuracy degradation caused by their entanglement. Finally, the false detection issue is greatly reduced through the implementation of adaptive geometric filters. The experimental results demonstrate that the proposed method achieves an average accuracy of 93.21% on rainy nights, indicating an average improvement of 12.7% over state-of-the-art methods. Additionally, without any parameter modifications, the proposed method is applicable to both normal and classic challenging scenes, such as nights, tunnels, rainy days, and shadows. The algorithm achieves an average accuracy of 96.2% and an average detection speed of 46 frames per second.

Connected and Autonomous Vehicles (CAVs) are emerging as a potential solution to merging safety problems. However, in mixed traffic scenarios where CAVs coexist with Human-driven Vehicles (HVs), challenges arise due to the lack of proactive cooperation and the limited length of the acceleration lane, complicating the merging processes of CAVs. In these cases, CAVs should actively seize the transit opportunity and perform safe and efficient merging. Failure to do so can lead to decreased traffic efficiency, increased fuel consumption and emissions, compromised self-merging capacities, and heightened crash risk. Therefore, this paper employs the roadside unit and proposes a two-level hierarchical self-merging strategy for CAVs to increase the merging efficiency while ensuring high safety. Since the Surrogate Safety Measures (SSM) can formulate reliable safety assessment and identify the merging conflict risk by setting appropriate threshold, the upper level uses a novel SSM-based method (i.e., the MInimum Acceleration Rate, MIAR) to determine the Merging Sequence (MS). A Theoretical Model for Merging Safety (TMMS) is developed to estimate the MIAR value, and the MIAR threshold is determined using Signal Detection Theory (SDT). At the lower level, the strategy recommends the optimized merging maneuvers, pre-generated using Sequential Quadratic Programming-Model Predictive Control (SQP-MPC), based on the determined MS and the CAV’s velocity. A case study at a real-world freeway merging area demonstrates the effectiveness of the MIAR in measuring merging conflict risk. Besides, numerous simulations are conducted, and the results demonstrate that the proposed strategy significantly improves merging success rate and overall traffic efficiency.

Driver gaze zone estimation (DGZE) is essential for detecting the driver’s state and taking over rule-making in intelligent driving systems. However, CNN-based multichannel models lack global feature extraction capability, with a large number of parameters and high computational complexity. Therefore, this paper proposes a novel method that uses 3-channel convolution-optimized ViT (3C-CoViT) to estimate the driver’s gaze zone. The method replaces the linear projection in the pure ViT structure with convolutional projection, converts the input images of different channels into image sequences, and then adds a convolutional feed-forward network to extract the local features of the markers, enhance the correlation of adjacent tokens in spatial dimensions, and improve the performance and efficiency of the model. We then pre-trained the model on the GazeCapture dataset based on transfer learning and then fine-tuned the model on the dataset built in the actual road experiment. To enhance the interpretability of the model, we presented a novel visualization method. Experimental results show that the proposed method can accurately identify driver gaze zones (98.04% average accuracy) and outperform state-of-the-art methods in terms of accuracy and reliability. Ablation studies proved the effectiveness of our proposed method over the pure ViT and the beneficial effects of transfer learning and 3-channel information input.

The difference between the desired trajectory of the drivers and the lane change trajectory of autonomous vehicles is one of the crucial factors that limit their development. This paper proposes a lane change trajectory prediction framework, referred to as LCT-DPP, which incorporates driver personality parameters to improve the acceptance of the lane change trajectories of autonomous vehicles. In order to identify the lane-changing styles of the drivers, a clustering algorithm Autoencoder (AE) -Gaussian Mixture Model (GMM) is first proposed to cluster the historical lane-changing trajectories and vehicle interaction parameters of the vehicles and classify the lane-changing styles of the driver into aggressive, normal, and conservative according to the clustering results. The encoder module and the fully connected layer are then used to predict the lane change style of the driver, and the output is a vector holding the lane change style features of the driver. Finally, the encoder-decoder model is developed for lane change trajectory prediction by fusing lane change style feature vectors and time series features and using them as inputs to the model. The proposed model is then validated through experiments. The obtained results show that the root mean square errors (RMSEs) of the TCN-LSTM-Attention based encoder-decoder model are respectively 0.312, 0.568, and 0.734 m for the time intervals of 1, 2, and 3 s, which indicates a very high prediction accuracy.

Identifying distracted drivers is crucial for enhancing driving safety and advancing intelligent driver assistance systems. Recently, researchers have applied Convolutional Neural Network (CNN) and Vision Transformer (ViT) models for driver state decision. However, both models often suffer from several issues such as numerous parameters and low detection efficiency. To address these challenges, this study proposes the Convolution Vision Transformer (CoViT) model for distracted driver identification, leveraging techniques such as Low Complexity Attention Mechanism (LCAM), Multi-scale Dilation Convolution (MSDC), and Depth Separable Convolution (DSC). Moreover, the CoViT model features a typical “pyramid” structure, enabling effective feature extraction across different scales. Subsequently, the proposed system is trained and evaluated using the publicly available driving behavior datasets SFD2 and 100-Driver, as well as real-world road experiments. Experimental results show that the CoViT model yields high recognition performance, with mean Accuracy (mAcc) scores of 95.17%, 97.89%, and 93.54% on the recorded dataset, SFD2 dataset, and 100-Driver dataset, respectively. These scores surpass those obtained by similar lightweight models. Furthermore, ablation experiments reveal that deep and dilated convolution significantly enhance model performance. In addition, the CoViT model demonstrates its applicability to real-time driving behavior detection tasks, with a parametric count of just 1.24M — a reduction of 2.67M compared to MobileNetV3 — and an online inference Frames Per Second (FPS) of 159.13.

The electronic differential control (EDC) system is an indispensable key technology for distributed-drive buses. However, compared to traditional buses with mechanical differentials, distributed-drive buses based on EDC have more severe tire wear and tire bias wear. To tackle these problems, we consider the reduction of tire wear, and propose an EDC strategy based on model predictive control (MPC) to improve the steering stability and extend the tire service life. First, nonlinear three-degree-of-freedom dynamics model and magic tire formula are established. Second, the slip ratio is selected as the observation of tire wear, and the underlying causes of tire wear are analyzed via simulation tests under different driving conditions. Third, a three-layer EDC strategy is designed. The upper layer is the EDC trigger condition and vehicle speed-following control; the middle layer is the calculation of control parameters such as the desired stability parameters, estimated sideslip angle, and slip state of the driving wheels; and the lower layer is the MPC-based driving torque controller to obtain the motor torque. Finally, the proposed MPC-based EDC strategy is validated and compared by simulation tests and real experiments. The comparative results demonstrate that the MPC-based EDC can achieve optimized differential torque control to effectively reduce the tire wear while improving the stability in different steering maneuvers.