Complexity controllable road network generation is pivotal for accelerated virtual simulation testing of autonomous vehicles (AVs), enabling the implementation of progressively challenging test scenarios through incrementally complex road networks, facilitating continuous evaluation of the target autonomous driving algorithms to expose functional flaws and performance boundaries. To this end, this study proposes a complexity controllable road network generation approach via optimized combination of realistic road elements. First, real-world urban road networks are decomposed into diverse road elements. Among these, elements with high potential collision risks (e.g., T-junctions, merging/diverging zones) are abstracted into parameter-configurable graph models defined by node and edge parameters. Then, the instantiation of these road element models is achieved using real-world cartographic data, followed by complexity assessment via a metric quantifying potential collision risk. Concurrently, a complexity evaluation function for road elements is formulated to assign complexity labels to each road instance. A complexity-tunable road network optimization model is developed. Taking the realism and compactness of the generated road networks as optimization objectives, the model selects road elements instance of varying complexity for optimal assembly in a non-intersecting and non-overlapping manner, yielding virtual road networks with controllable complexity while preserving real-world characteristics. Finally, experimental validation was conducted using Xi’an cartographic data, generating 4,500 virtual road networks across 15 distinct complexity levels through combination of road elements with different complexity. To validate the effectiveness of the generated road networks, comparative virtual simulation experiments are conducted on both the generated networks and real-world road networks. Experimental results demonstrate that: (1) The mean relative error between the geometric parameters of road elements in the generated networks and the expected values derived from real-world cartographic data clustering is less than 1.5%, confirming superior environmental reconstruction fidelity; (2) Lane-change test scenarios constructed using the generated road networks successfully identified performance limits of the target autonomous driving algorithm and functional deficiencies under lateral approach conditions; (3) The complexity of the generated road networks shows a strong correlation (absolute correlation coefficients exceeding 0.84) with the uncomfortable driving duration and average vehicle speed in continuous free-driving scenarios, which confirms that higher-complexity road networks correspond to more rigorous testing challenges; (4) Compared to real-world road networks, the generated road networks achieve 32.4% average mileage compression when traversing an equivalent number of road elements, significantly enhancing testing efficiency. 

Proving ground testing has become a standard methodology for the development and validation of autonomous vehicles in the automotive industry. However, it suffers from inherent limitations in efficiency, cost, and scenario coverage. Combined virtual-real testing (CVRT) offers a promising alternative by integrating virtual scenarios with physical vehicles and the environment, enhancing scenario coverage and test flexibility. Nevertheless, few studies have systematically investigated its effectiveness and applicability. To address this gap, this study develops a digital-twin-based CVRT system and conducts consistency verification experiments, taking the autonomous emergency braking (AEB) system test as a case study. Four typical scenarios selected from C-NCAP (China New Car Assessment Programme) 2024 were tested at speeds of 30, 40, and 50 km/h, utilizing both real-world and CVRT methods, with each experiment repeated 15 times. Vehicle dynamics data were collected, and the Fréchet distance metric was used to quantify similarity, whereas statistical hypothesis testing was used to assess differences in time-to-collision (TTC) trigger times. The results show that the average Fréchet distance ratio between the CVRT and real-world tests almost approaches 1.0, and the differences in the TTC trigger times were not statistically significant. However, the results of the simulation experiments differed significantly from those of the real-world tests (0.528 m/s in speed and 1.150 m/s² in acceleration higher than the CVRT). Additionally, the data communication delay between the CVRT platform and the physical autonomous vehicle under test remained well below tolerable thresholds. These results indicate high consistency between CVRT and real-world testing. Furthermore, CVRT achieved considerable improvements in testing efficiency, saving approximately 40%–70% compared with real-world testing.

A widely studied and concerned problem in the traffic network research is how to optimize the location and number of road traffic detectors, so as to obtain real-time and accurate diversified traffic situation information and provide a comprehensive information basis for traffic control departments and as a basis for reasonable decision-making. The key to this problem is to select a suitable detector type and build a decision model according to the research purpose. At the same time, considering the constraints such as the investment cost limit and the number of road sections, appropriate heuristic algorithm should be used to solve the model to get the best number and location of detectors. This paper summarized the optimal layout of road traffic detectors from the types of road traffic detectors, application scenarios, data acquisition indexes and research objectives of various optimization layout studies. Firstly, the detector was divided into two categories according to the installation mode: stationary traffic detector and mobile detector, and the principle, characteristics, advantages and disadvantages of each type of detector were described in detail. Secondly, the application of various types of road traffic detectors in different scenarios and the corresponding data acquisition indicators are given. Then, according to the research purpose of optimization layout methods in the research literature, the optimization layout problems of road traffic detectors were divided into three types: user-oriented travel time estimation, traffic flow observation/estimation, and traffic event detection. And this paper discussed the development course, development direction, problem research model constructed, problem solving methods, and existing shortcomings of these studies. Finally, it summarized a large number of existing studies. And it pointed out that in the complex situation of large traffic network scale, prominent traffic uncertainty and rapid development of wisdom, future research should take the diversity of traffic information detection as the leading factor, fully consider the combination arrangement of different types of traffic detectors, various uncertainties in the traffic network and various scenarios, etc., so as to build a complete optimization model to solve the optimization arrangement of road traffic detectors.

This research addresses the challenging task of classifying drivers’ emotions to increase their awareness of their driving behaviors. It recognizes the common issue of driver emotions, which often leads to the neglect of poor driving practices. By automatically detecting and identifying these behaviors, drivers can proactively obtain valuable insights to reduce potential accidents. This study proposes a comprehensive facial recognition model for drivers that uses a unified architecture comprising a convolutional neural network (CNN), a recurrent neural network (RNN), and a multilayer perceptron (MLP) classification model. Initially, a faster region-based convolutional neural network (R-CNN) was employed for accurate and efficient facial detection of drivers in live and recorded videos. Features are extracted from three CNN models and merged via advanced techniques to create an ensemble classification model. Moreover, the improved Faster R-CNN feature learning module is replaced with a new convolutional neural network module, VGG16, which maximizes the precision and effectiveness of facial detection in our system. Significant accuracy results of 89.2%, 97.20%, 99.01%, 93.65%, and 98.61% are shown in evaluations of our suggested facial detection and facial expression recognition (DFER) datasets, including the EMOTIC, CK+, FERPLUS, AffectNet, and custom datasets. These datasets were meticulously acquired in a simulated environment, necessitating the creation of several custom datasets. This research highlights the potential of deep ensemble classification in improving driver emotion recognition, thereby contributing to enhanced road safety.

The event wherein vehicles from the adjacent lane execute a lane-change maneuver, cutting into the lane occupied by autonomous vehicles, epitomizes a typical high-risk scenario on expressways within the domain of autonomous driving. Replicating such scenarios for testing on actual expressways involves significant safety risks. Virtual simulation test is one of the best approaches to addressing this issue. In order to automatically generate mass high-fidelity expressway lane-change cut-in test scenarios, this paper presents an automatic generation algorithm of lane-change cut-in test scenarios for autonomous vehicles based on TsGAN (Time-Series Generative Adversarial Network). In this algorithm, the time headway and lateral gap at the cut-in moment are taken as the evaluation metrics for scenario risk assessment, and 2853 instances of lane-change cut-in scenarios with four different risk levels are extracted from the real expressway trajectory dataset highD. A model for generating lanechange cut-in test scenarios is established based on TsGAN, and is trained with the extracted real trajectory data. Then, the model is employed to generate the trajectories for the lane-change vehicle and the tested vehicle before the cut-in moment. To authenticate the generated trajectories, the distribution similarity and spectral error between the generated and the real trajectories are compared. Furthermore, an in-depth analysis of the kinematic interplay between the two vehicles at the cut-in moment and the distribution of trajectory parameters in the generated scenarios is performed to validate the coverage of the generated scenarios within naturalistic settings. The findings can be summarized as follows: (1) as illustrated by the distribution of key trajectory parameters, the average similarity between the generated and the real lane-change trajectories is 79.7%, with an average spectral error less than 8%, and more than 83.2% of the generated trajectories are within the buffer of the most analogous real trajectories, indicating a notable fidelity of the generated trajectories; (2) as compared with the collected real scenarios, the generated instances exhibit a more expansive coverage and a more even distribution within parameter intervals, and the time headway and lateral gap between the lane-change vehicle and the tested vehicle at the cut-in moment decrease by 17.83% and 16.37% in average, respectively, the distribution range of trajectory parameters expands by 19.44%, signifying a heightened coverage of the generated scenarios; and (3) the proposed TsGAN-based generation model has the capability of emulating lane-change cut-in test scenarios with four different risk levels, exhibiting pronounced specificity.

Vehicle Re-identification (Re-ID) has drawn extensive exploration recently; nevertheless, the issue of accurately distinguishing features in latent space across varying vehicle poses, remains a challenging hurdle for real-world application of Vehicle Re-ID. To address this challenge, we supply a novel idea which projects the various-pose vehicle images into a unified target pose so as to promote the discriminative capability of vehicle Re-ID model. Acknowledging the labor and cost of paired data for the same vehicle images across different traffic surveillance cameras in practical scenarios, we propose the pioneering Pair-flexible Pose Guided Image Synthesis for vehicle Re-ID, denominated as VehicleGAN. Our method is adept at both supervised (paired images of same vehicle) and unsupervised (unpaired images of any vehicle) settings, and bypasses the need of geometric 3D model information. Furthermore, we propose a novel Joint Metric Learning (JML) method to facilitate the effective fusion of both real and synthetic data. Comprehensive experimental analyses conducted on the public VeRi-776 and VehicleID datasets substantiate the precision and efficacy of our proposed VehicleGAN and JML.

Deep learning-based intelligent vehicle perception has been developing prominently in recent years to provide a reliable source for motion planning and decision making in autonomous driving. A large number of powerful deep learning-based methods can achieve excellent performance in solving various perception problems of autonomous driving. However, these deep learning methods still have several limitations, for example, the assumption that lab-training (source domain) and real-testing (target domain) data follow the same feature distribution may not be practical in the real world. There is often a dramatic domain gap between them in many real-world cases. As a solution to this challenge, deep transfer learning can handle situations excellently by transferring the knowledge from one domain to another. Deep transfer learning aims to improve task performance in a new domain by leveraging the knowledge of similar tasks learned in another domain before. Nevertheless, there are currently no survey papers on the topic of deep transfer learning for intelligent vehicle perception. To the best of our knowledge, this paper represents the first comprehensive survey on the topic of the deep transfer learning for intelligent vehicle perception. This paper discusses the domain gaps related to the differences of sensor, data, and model for the intelligent vehicle perception. The recent applications, challenges, future researches in intelligent vehicle perception are also explored.