For trajectory prediction within autonomous vehicle planning and control, conditional variational autoencoders (CVAEs) have shown promise in accurate and diverse modeling of agent behaviors. Besides accuracy, explainability is also crucial for the safety and acceptance of learning-based autonomous systems, especially in autonomous driving. However, the latent distributions learned by CVAE models are often implicit and thus possess low explainability. To address this, we propose a semi-supervised generative modeling framework, \textbf{\textit{PrefCVAE}}, which utilizes partially and weakly labelled preference pairs to imbue the CVAE's latent representation with semantic meaning. This approach enables the system to estimate measurable attributes of the agents, and to generate manipulable predictions under the CVAE framework. Results show that incorporating our preference loss allows a CVAE-based model to make conditional predictions using the semantic factor of prediction average velocity. Our augmented framework also does not significantly degrade the baseline accuracy of prediction. Additionally, we show that the latent values learned with PrefCVAE better represent the semantic information contained in the data. Finally, we discuss the potential of this loss design to extend to other machine learning applications beyond trajectory prediction, as well as essential tricks for adaptation of human labeling. We hope that our empirical study offers the broader representation learning community a fresh perspective on inductive bias for disentangled and explainable latent representations in deep generative models. Specifically, we demonstrate that preference pair supervision, a simple and cost-effective approach, can effectively aid in learning semantic meanings for sampling-based generative models like the CVAE.

Connected and Automated Vehicles (CAVs) have the poten- tial to revolutionize transportation systems, but their integration with human-driven vehicles (HDVs) in mixed traffic environments presents significant challenges, particularly in complex scenarios such as on-ramp merging. This paper addresses the challenge of on-ramp merging for CAVs in mixed traffic environments, proposing a novel approach called QMIX-QLambdaM. We formulate the problem as a Centralized Train- ing with Decentralized Execution (CTDE) Cooperative Multi-Agent Re- inforcement Learning (MARL) task, capable of handling dynamic sce- narios with both CAVs and HDVs. QMIX-QLambdaM enhances the QMIX algorithm by incorporating Q(λ) returns for improved value es- timation and an action masking mechanism for safer action selection. Our comprehensive experiments demonstrate that QMIX-QLambdaM consistently outperforms state-of-the-art algorithms, including QMIX, MAA2C, and COMA, across various performance metrics related to traf- fic efficiency and safety. The proposed method exhibits superior adapt- ability across different traffic densities, maintaining high performance in terms of safety, efficiency, and overall rewards. Furthermore, case stud- ies illustrate QMIX-QLambdaM’s ability to generate effective strate- gic control for both main-lane and merging-lane vehicles, showcasing smoother driving behavior and better collision avoidance compared to baseline methods. The learning curve comparison also reveals QMIX- QLambdaM’s advantage in credit assignment compared to other CTDE baselines for the formulated problem. The code are available at https: //github.com/ayton-zhang/MARL_qmix_merging.

Road traffic safety is a pressing global concern, with millions of yearly fatalities and injuries. This study aims to address the detection of abnormal driving behaviour. Traditional supervised approaches face limitations due to the need for labelled abnormal driving data. To overcome this challenge, semi-supervised machine learning models are explored and developed in this research.
Machine learning is utilized for abnormal driving behaviour detection because it offers a data-driven approach that adapts to different scenarios and captures subtle patterns. Furthermore, its scalability allows for efficient analysis of large datasets, leading to accurate identification of abnormal driving behaviour and valuable insights for enhancing road safety measures. Most existing machine learning (ML) based abnormal driving detectors rely on (fully) supervised ML methods, which require substantial labelled data. However, in the real world, labels are only sometimes available, and labelling large amounts of data is tedious. Thus, there is a need to employ unsupervised or semi-supervised methods to make the detection process more feasible and efficient. Luckily, it is possible with the advent of deep neural networks, especially autoencoder-based ones. This thesis develops and compares three ML methods: supervised (e.g. XGBoost and Random Forest), unsupervised ML (e.g. Isolation Forest and Robust Covariance), and semi-supervised ML (Hierarchical Extreme Learning Machines). Comparison results show that the semi-supervised deep learning model outperforms unsupervised methods exhibiting higher prediction accuracy and delivering acceptable results compared to the fully supervised models.
Moreover, previous ML-based approaches predominantly utilize basic car motion features (such as velocity and acceleration) to label and predict abnormal driving behaviours. In contrast, this thesis introduces Surrogate Measures of Safety (SMOS) as features for ML models to identify abnormal driving behaviour.
The results indicate that the supervised model performs best under the same conditions. However, relying on a large amount of labelled data in supervised models can pose challenges in real-life scenarios or when dealing with massive datasets. The study highlights the significance of Surrogate Measures of Safety (SMOS) and demonstrates the potential of HELM in effectively identifying abnormal driving behaviour. The introduction of SMOS significantly improves the performance of both unsupervised and semi-supervised models. The unsupervised model shows the most substantial improvement, increasing accuracy from less than 50% to over 90%.
While the Isolation Forest and Robust Covariance models fail to detect abnormal driving behaviour without including SMOS, the semi-supervised HELM model exhibits promising results even without SMOS. However, further research is necessary to address limitations and enhance the findings. While valuable, the current dataset used in this study may only encompass some types of abnormal driving behaviour. Future research should incorporate a more diverse dataset that covers a broader range of abnormal driving behaviours. The analysis should include multiple SMOS features, such as Post Encroachment Time (PET), to comprehensively understand abnormal driving behaviour and improve safety measures.

Lane-changing behaviour detection is a critical aspect of driving safety and traffic management. This study focuses on detecting sudden lane changes as a subset of abnormal driving behaviours. By analyzing the characteristics of abrupt lane changes, the aim is to develop effective data-driven unsupervised machine learning (ML) methods for their detection and classification. Three unsupervised ML models, namely Isolation Forest, Local Outlier Factor, and Robust Covariance are evaluated and compared using a dataset of lane-change events. The results show that the Isolation Forest and Local Outlier Factor models outperform the Robust Covariance model, with the Local Outlier Factor model excelling in precision and overall accuracy, achieving the best overall detection rate. Both Robust Covariance and Isolation Forest deliver satisfactory results. Conversely, the Robust Covariance model exhibits poor performance. The findings verify the capability of data-driven ML methods for enhancing road safety and driving experiences through effective detection of sudden lane changes using vehicle motion information data. Future work involves further improving the accuracy and reliability of the ML models, validating their generalizability on larger datasets, incorporating contextual information, and exploring their real-time implementation in driving assistance systems.

Advanced Driver Assistance Systems (ADAS) are becoming more available and will become mandatory for all new vehicle models from 2022 onward. In order to achieve the highest safety benefits, it is important that these systems are available. Lane Keep Assist (LKA) is part of ADAS and assists the driver in the lateral control of the vehicle. Lane markings are used by both human drivers and machine vision to stay on the road, but factors contributing to lane marking detection in different driving conditions are mostly unknown. A field test was conducted on Dutch provincial roads to evaluate lane marking visibility properties in relation to the LKA detection performance of different sensor types. The LKA detection performance of the mono camera was found to be higher in most weather and illumination conditions than the detection performance of the mono camera with infrared. The mono camera with infrared had a higher detection performance during rain in nighttime conditions than during dry daytime conditions. The highest detection performance for the mono camera and the mono camera with infrared were 97% in dry nighttime conditions and 91,4% in sunset conditions, respectively. Binary logistic regression was used to determine the effect of lane marking properties on the lane detection performance. A profiled lane marking type was found to increase the detection likelihood by 6-8 times as opposed to a smooth lane marking type. Other visibility properties, such as retroreflectivity and contrast with the road surface, were not found to be a significant contributor to the detection performance.