Connected and automated driving confronts critical complex traffic scenarios and safety, which attracts increasing attention from academia and industry. To enhance the intelligent safety for connected and automated vehicles (CAVs), including functional safety, safety of the intended functionality (SOTIF), and cybersecurity, efforts are required to seek solutions from fundamental theory, guarantee framework and general/systematic point of view. This feature topic aims to provide a platform for academia, industry, and policymakers, on which they could present the latest research and engineering experience in developing and applying novel technologies. The topics cover SOTIF, cybersecurity, artificial intelligence for CAV safety, risk assessment-based vehicle path planning and decision-making under uncertain environments.

The safety of the intended functionality (SOTIF) has become one of the hottest topics in the field of autonomous driving. However, no testing and evaluating system for SOTIF performance has been proposed yet. Therefore, this paper proposes a framework based on the advanced You Only Look Once (YOLO) algorithm and the mean Average Precision (mAP) method to evaluate the object detection performance of the camera under SOTIF-related scenarios. First, a dataset is established, which contains road images with extreme weather and adverse lighting conditions. Second, the Monte Carlo dropout (MCD) method is used to analyze the uncertainty of the algorithm and draw the uncertainty region of the predicted bounding box. Then, the confidence of the algorithm is calibrated based on uncertainty results so that the average confidence after calibration can better reflect the real accuracy. The uncertainty results and the calibrated confidence are proposed to be used for online risk identification. Finally, the confusion matrix is extended according to the several possible mistakes that the object detection algorithm may make, and then the mAP is calculated as an index for offline evaluation and comparison. This paper offers suggestions to apply the MCD method to complex object detection algorithms and to find the relationship between the uncertainty and the confidence of the algorithm. The experimental results verified by specific SOTIF scenarios proof the feasibility and effectiveness of the proposed uncertainty acquisition approach for object detection algorithm, which provides potential practical implementation chance to address perceptual related SOTIF risk for autonomous vehicles.

The increasingly intelligent and connected vehicles have brought many unprecedented automotive cybersecurity threats, which may cause privacy breaches, personal injuries, and even national security issues. Before providing effective security solutions, a comprehensive risk assessment of the automotive cybersecurity must be carried out. A systematic cybersecurity risk assessment framework for automobiles is proposed in this study. It consists of an assessment process and systematic assessment methods considering the changes of threat environment, evaluation target, and available information in vehicle lifecycle. In the process of risk identification and risk analysis, the impact level and attack feasibility level are assessed based on the STRIDE model and attack tree method. An automotive cybersecurity risk matrix using a global rating algorithm is then constructed to create a quantitative risk metric. Finally, the applicability and feasibility of the proposed risk assessment framework are demonstrated through a use case, and the results prove that the proposed framework is effective. The proposed assessment framework helps to systematically derive automotive cybersecurity requirements.

As intelligent vehicles become increasingly computerized and networked, they gain more autonomous capabilities. However, they are also becoming more exposed to cyber-threats which are likely to be a more prominent concern. This paper proposes a cyber-attack detection method for autonomous vehicles based on secure estimation of vehicle states, with an example application under attacks in the vehicle localization system. To investigate the effects of vehicle model and estimator on the attack detection performance, different nonlinear vehicle dynamic models and estimation approaches are employed. The deviation between the measurement from the onboard sensors and the state estimation is monitored in real time. With the designed vehicle state estimator and preset threshold, the cyber-attack detection algorithm is further developed for autonomous vehicles, whose performance is tested in simulations where the vehicle localization system is assumed to be compromised during a double lane change maneuver. The test results demonstrate the feasibility and effectiveness of the proposed cyber-attack algorithm. In addition, the results illustrate the impacts of vehicle nonlinear characteristics on the cyber-attack detection performance. Beyond this, the effects of different vehicle models on the attack detection performance, as well as the selection of suitable filtering approaches for the attack detection, are also discussed.

Before highly automated vehicles (HAVs) become part of everyday traffic, their safety has to be proven. The use of human performance as a benchmark represents a promising approach, but appropriate methods to quantify and compare human and HAV performance are rare. By adapting the method of constant stimuli, a scenario-based approach to quantify the limit of (human) performance is developed. The method is applied to a driving simulator study, in which participants are repeatedly confronted with a cut-in manoeuvre on a highway. By systematically manipulating the criticality of the manoeuvre in terms of time to collision, humans’ collision avoidance performance is measured. The limit of human performance is then identified by means of logistic regression. The calculated regression curve and its inflection point can be used for direct comparison of human and HAV performance. Accordingly, the presented approach represents one means by which HAVs’ safety performance could be proven.

Path-following control is one of the key technologies of autonomous vehicles, but the complex coupling effects and system uncertainties of vehicles can degrade their control performance. Accordingly, this study proposes targeted methods to solve different types of coupling in vehicle dynamics. First, the types of coupling are figured out and different handling strategies are proposed for each type, among which the coupling caused by steering angle, unsaturated tire forces, and load transfer can be treated as uncertainties in a unified form, such that the coupling effects can be treated in a decoupling way. Then, robust control methods for both lateral and longitudinal dynamics are proposed to deal with the uncertainties in dynamic and physical parameters. In lateral control, a robust feedback–feedforward scheme is utilized in lateral control to deal with such uncertainties. In longitudinal control, a radial basis function neural network-based adaptive sliding mode controller is introduced to deal with uncertainties and disturbances. In addition, the tire saturation coupling that cannot be handled by controllers is treated by a proposed speed profile. Simulation results based on the CarSim–Simulink joint platform evaluate the effectiveness and robustness of the proposed control method. The results show that compared with a well-designed robust controller, the velocity tracking performance, lateral tracking performance, and heading tracking performance improve by 55.68%, 34.26%, and 52.41%, respectively, in the double-lane change maneuver, and increase by 87.79%, 30.18%, and 9.68%, respectively, in the ramp maneuver.

Faults that develop in vehicle sensors have the potential to propagate unchecked throughout control systems if undetected. Automatic fault diagnosis and health monitoring algorithms will become necessary as automotive applications become more autonomous. The current fault diagnosis systems are not effective for complex systems such as autonomous cars where the case of simultaneous faults in different sensors is highly possible. Therefore, this paper proposes a novel fault detection, isolation and identification architecture for multi-fault in multi-sensor systems with an efficient computational burden for real-time implementation. Support Vector Machine techniques are used to detect and identify faults in sensors for autonomous vehicle control systems. In addition, to identify degrading performance in a sensor and predict the time at which a fault will occur, a novel predictive algorithm is proposed. The effectiveness and accuracy of the architecture in detecting and identifying multiple faults as well as the accuracy of the proposed predictive fault detection algorithm are verified through a MATLAB/IPG CarMaker co-simulation platform. The results present detection and identification accuracies of 94.94% and 97.01%, respectively, as well as a prediction accuracy of 75.35%.

Research on vehicle travel destinations mostly only consider vehicle trajectory data and ignore the influence of other multi-source data, such as weather, time, and points of interest (POI). This study proposes a destination prediction method based on multi-source data, and a multi-input neural network model is established. In terms of the coding of vehicle trajectory data, a GeoHash to vector (Geo2vec) model is proposed to realize the characterization of the trajectory. As for the coding of temporal features, a cyclic coding model is proposed based on trigonometric functions. For the coding of POI, an origin–destination POI matrix (OD-POI) model is proposed based on the state transition probability. Experimental results show that in terms of the average distance and root-mean-square distance deviations, Geo2vec reveals reductions of 4.51% and 5.63% compared to word to vector (Word2vec), and cyclic encoding shows reductions of 6.35% and 6.67% compared to label encoding; further, the method of OD-POI state transition probability is reduced by 5.85% and 6.4%, and the model based on multi-source data is 17.29% and 17.65% lower than the model based on trajectory data only. Finally, the cyclic encoding is reduced by 48.60% in the data dimension compared to one-hot encoding. Accurate destination prediction will help improve the efficiency of automotive human–computer interaction

Recent years have seen the rapid development of autonomous driving systems, which are typically designed in a hierarchical architecture or an end-to-end architecture. The hierarchical architecture is always complicated and hard to design, while the end-to-end architecture is more promising due to its simple structure. This paper puts forward an end-to-end autonomous driving method through a deep reinforcement learning algorithm Dueling Double Deep Q-Network, making it possible for the vehicle to learn end-to-end driving by itself. This paper firstly proposes an architecture for the end-to-end lane-keeping task. Unlike the traditional image-only state space, the presented state space is composed of both camera images and vehicle motion information. Then corresponding dueling neural network structure is introduced, which reduces the variance and improves sampling efficiency. Thirdly, the proposed method is applied to The Open Racing Car Simulator (TORCS) to demonstrate its great performance, where it surpasses human drivers. Finally, the saliency map of the neural network is visualized, which indicates the trained network drives by observing the lane lines. A video for the presented work is available online, https://youtu.be/76ciJmIHMD8 or https://v.youku.com/v_show/id_XNDM4ODc0MTM4NA==.html.

In pedestrian-to-vehicle collision accidents, adapting safety measures ahead of time based on actual pose of pedestrians is one of the core objectives for integrated safety. It can significantly enhance the performance of passive safety system when active safety maneuvers fail to avoid accidents. This study proposes a deep learning model to estimate 3D pose of pedestrians from images. Since conventional pedestrian image datasets do not have available pose features to work with, a computer graphics-based (CG) framework is established to train the system with synthetic images. Biofidelic 3D meshes of standing males are first transformed into several walking poses, and then rendered as images from multiple view angles. Subsequently, a matrix of 50 anthropometries, 10 gaits and 12 views is built, in total of 6000 images. A two-branch convolutional neural network (CNN) was trained on the synthetic dataset. The model can simultaneously predict 16 joint landmarks and 14 joint angles of pedestrian for each image with high accuracy. Mean errors of the predictions are 0.54 pixels and ??0.06°, respectively. Any specific pose can then be completely reconstructed from the outputs. Overall, the current study has established a CG-based pipeline to generate photorealistic images with desired features for the training; it demonstrates the feasibility of leveraging CNN to estimate the pose of a walking pedestrian from synthesized images. The proposed framework provides a starting point for vehicles to infer pedestrian poses and then adapt protection measures accordingly for imminent impact to minimize pedestrian injuries.