Automated vehicles (AVs) represent a significant leap forward in transportation, aiming to enhance road safety, increase comfort, and improve traffic efficiency. As technology progresses to SAE Level 3 and higher, drivers are increasingly able to engage in non-driving-related activities. However, this (r)evolution raises challenges concerning driver's perceived risk and trust in AVs, which are crucial factors influencing the acceptance of AVs. This dissertation aims to enhance perceived safety and trust in AVs through experimental studies, computational modelling and user interfaces (UIs) design.

The initial phase of this dissertation focused on how drivers' perceived risk and trust when using AVs evolve in close encounters with other road users. We developed regression-based perceived risk and trust models based on a simulator study with 25 participants involving merging and hard braking scenarios on motorways. The proposed models reveal that perceived risk is dynamically influenced by driving conditions and sensitive to individual factors such as driving experience and gender with experienced and male drivers generally perceiving lower risk. Notably, a decrease in trust after high-risk encounters was observed, indicating a close relationship between perceived risk and trust in AVs. Additionally, physiological responses were observed as potential indicators of perceived risk in critical driving scenarios.

To develop a tool for gaining insights on perceived risk, we put forward a novel computational model called potential collision avoidance difficulty (PCAD) model. Drawing inspiration from Fuller's Risk Allostasis Theory and the looming phenomenon, PCAD evaluates the difficulty of avoiding potential collisions by calculating minimal control effort through braking or/and steering needed to navigate safely. By integrating visual looming, factors in the uncertain behaviour of surrounding vehicles, control inaccuracies of the subject vehicle, and potential collision severity, PCAD provided an accurate population-level fitting of perceived risk in our own dataset on highway merging and a published dataset on obstacle avoidance. The findings highlight the need to account for both the longitudinal and lateral dimensions of driving condition, and uncertain behaviours of surrounding vehicles when interpreting perceived risk.

Further exploration of perceived risk was achieved through the creation of a large-scale dataset of perceived risk using an online survey. This new dataset provided time-continuous perceived risk in dynamic driving conditions. A total of 105 events was created including merging, hard braking and lane changes on motorways, while systematically varying multiple control parameters (such as relative speed and distance) to achieve different levels of event criticalities. Deep neural networks (DNNs) were then trained on this dataset to fit perceived risk, and SHapley Additive exPlanations (SHAP) was used to identify the key contributors to perceived risk in the continuous time domain. Aligned with the PCAD model developed previously, the results highlighted the importance of the relative motion information, particularly the distance to other road users and the uncertainty of surrounding vehicle behaviour in shaping perceived risk. This approach not only discerns the dynamics of perceived risk by systematically analysing interactions with other road users but also provides a guide for future modelling of perceived risk. The development of this extensive dataset fills the gap by providing the lacking continuous perceived risk data, thereby supporting further research on perceived risk.

The last contribution of this dissertation was on enhancing perceived safety and trust through optimised design of UIs. A simulator experiment demonstrated that multimodal UIs incorporating both visual and auditory modalities enhanced perceived safety and trust the most. Manoeuvre information delivered through the auditory modality was particularly effective in enhancing trust and acceptance. The findings indicate the benefits of the UIs in enhancing perceived safety and trust but also showed the limitations of using UIs alone during highly critical events. This part of the work suggests that the design of UIs for partially automated vehicles shall include automation information via visual and auditory modalities to enhance perceived safety and trust.

This dissertation makes several contributions to the field of perceived risk research in AVs. First, it provides foundational insights into perceived risk, demonstrating the significant influences of driving conditions, manoeuvre uncertainties and individual personal characteristics. The computational perceived risk models demonstrate strong predictive power in perceived risk and offer a deep understanding of how perceived risk is shaped in dynamic driving conditions. Additionally, the rich dataset obtained in this dissertation, which includes event-based discrete data and time-continuous data on perceived risk, serves as a new and open resource for future perceived risk research. Lastly, the practical evaluation of the design of UI provided actionable recommendations in enhancing trust and perceived safety, particularly through manoeuvre information delivered using auditory modality. These contributions advance the understanding, modelling, and practical application of perceived risk in automated driving environments, supporting the broader acceptance and integration of AVs.

The dissertation presents various opportunities for the advancement of AV technology and its integration with human factors. Building on the comprehensive datasets, computational models and insights gained in this dissertation, future studies should focus on further refining computational models to capture perceived risk in general scenarios. Expanding data collection efforts to include on-road tests, and more diverse participants will also enhance the generalisability of the findings. Additionally, the design of adaptive UIs that fit individual preferences remains a promising direction for future research.@en