The level of automation in vehicles is growing. But until all vehicles are completely automated, there will be a transition period where automated vehicles and human drivers coexist. Because these road users will coexist, it is necessary that automated vehicles understand human drivers and vice versa. This study aims to create a model that predicts human risk perception in different driving scenarios, to provide an understanding of the fundamental features of human threat perception while driving. The model created is a multi-criteria decision-making process that uses KITTI Vision Benchmark data as an input. This model is tested against the data gathered by an online survey, where 1918 participants answered the question: "How high is the risk on a scale from 0-10?" for 100 situations, chosen from the KITTI Vision Benchmark data. The survey response data is then compared to the model. Analysis of the survey data revealed that risk perception of driving situations is non-linear in the extremities of risk, showing that the input image are perceived as normally distributed instead of uniformly distributed. The comparison further shows that a model with features and weights solely based on literature is only slightly capable of predicting the risk of situations with a Pearson correlation coefficient with the survey responses of 0.28, whereas a model with feature weights optimised is moderately capable of predicting the risk of situations with a correlation coefficient of 0.57. However, multivariate regression is better capable of predicting risk with a correlation coefficient of 0.70, and shows that features and weights based on literature were not enough to establish an accurate model. The features that have the most impact on the result are the information about other road users' location and heading, the ego vehicle velocity, and the road type. Further research should focus on determining if speed is the cause of the perceived risk, or if it follows from other unknown predictors. Furthermore, an extension of the questions asked of the participants and the usage of videos instead of images can clarify the discrepancy between the literature-based model and the perceived risk of the participants.

Naturalistic driving research with a focus on trucks has been gaining momentum in the past decade. With the advancement in sensor technology and access to big data, it becomes possible to understand driver behaviour at a more fundamental level. This can assist in mitigating the impact trucks have on the environment while enhancing safety. Several studies have worked towards examining the predictability of driving behaviour through driver profiling (i.e. scoring a driver's behaviour or classifying drivers by assigning them different categories such as aggressive/non-aggressive). However, little research still focuses on the importance and impact of individual features used to develop these models. In the current study, an analysis of driving data from 1,727 trucks recorded over one year as part of a Dutch Field Operational Test (FOT) has been performed. This FOT, to date, has not been investigated in the published academic literature. Recent studies have analysed historical location data to assess risk associated with specific routes and environments. This is being used to provide notifications to drivers around work zones to mitigate the impact of accidents. The current thesis extends the geo-specific analysis of (truck) driving data by analysing stability in truck driving behaviour with a focus on time and location (urban areas and motorways). Here correlation analysis has been used to explore stability. Correlational analysis elucidates that metrics such as the number of headway warnings, braking events and lane departure warnings are stable over space and time. A discussion reflects on the role of vehicle characteristics (i.e. mass and engine power) towards stability. It is concluded that in the case of spatial stability: Mean point speed has higher stability on motorways than in urban areas. It has been determined that trucks with higher mass and lower engine power tend to have lower mean speed than the norm. Contrary to mean point speed, headway warnings show higher stability in urban areas than on motorways. Braking events and lane departure warnings exhibit high stability. Secondly, a strong correlation between (t and t+1) hours over the entire day is observed for temporal stability. This research is a precursor to building generalised models for profiling drivers and assessing various driving patterns. An in-depth understanding of different driving patterns can help driver coaching companies better understand metrics when time and location are factored in before providing targeted feedback. Apart from this can also facilitate fleet management. The code for analysing this dataset is accessible online and may stimulate future researchers to explore this dataset further.

Pedestrians today are very vulnerable on urban roads. Clear communication between drivers and pedestrians is one way to reduce their plight. Non-verbal communication in particular plays an important role in road safety, and eye contact is a kind of non-verbal communication that has the potential to minimize on-road collisions. However, with the advent of automated vehicles, driver-pedestrian eye contact loses its meaning since there is no longer a driver. It is therefore useful to study and detect eye contact so that the knowledge obtained may be applied to automated vehicles of the future. To this end, the following research goals were adopted: (a) What is eye contact between a pedestrian and a driver in a car? How can eye contact be defined/operationalized using an algorithm?, (b) How accurate is the algorithm that operationalizes eye contact?, and (c) How is it possible to use two eye-trackers with inertial measurement units (IMUs) and pedestrian recognition in a Toyota Prius car to reconstruct the entire driver-pedestrian interaction through a 3-D animation? An indoor experiment, designed to resemble a driver-pedestrian interaction at a pedestrian crossing was conducted with 31 participants. Participants’ (pedestrians’) eyes were tracked using a Tobii Pro Glasses 2 eye-tracker and the researcher’s (driver’s) eyes were tracked using a Smart Eye Pro dx eye-tracker, both of which were synchronized. Participants’ locations were also tracked using a stereo camera equipped with pedestrian detection capabilities. Pedestrians imagined that they were on a real road and performed six types of trials where they stood on / crossed from the left / right side curb in front of the stationary vehicle while either making eye contact or not making eye contact with the driver. The order of the trials was randomized, and each trial consisted of 3 repetitions of a driver-pedestrian interaction. If the driver and pedestrian were looking at each other at the same time there was eye contact, otherwise there was no eye contact. Significant differences in the percentages of eye contact between pedestrians standing on the left (median duration of 0.42 s) and the right (median duration of 0.54 s) were found. No significant differences in the percentages of eye contact between pedestrians crossing from the left (median duration of 1.23 s) and the right (median duration of 1.39 s) were found. Eye contact instants within trials were algorithmically detected by finding the angle between the 3-D gaze direction vectors of the driver and the pedestrian, and comparing it to an ‘eye contact threshold’. Trials were classified as either involving eye contact or not involving eye contact based on their percentages of eye contact instants. The classification performance of the algorithm was quantified using two ground truths: (1) Imposed eye contact (in half of the trials, participants were instructed to make eye contact; in the other half, participants were instructed not to make eye contact), and (2) Manually annotated areas of interest (AOIs) from the Tobii Pro Glasses 2 showing pedestrian eye contact seeking. The algorithm’s performance was found to be fair/poor and eye contact could be detected with an accuracy of 15-30°. A 3-D reconstruction of the driver-pedestrian interaction was achieved (in the form of an animation) by using the locations, head orientations and gaze directions of the driver and the pedestrian. This thesis provides objective measurements of driver-pedestrian eye contact and demonstrates how eye contact may be detected and reconstructed for use in automated vehicles of the future.

Highly automated vehicles may lead to vehicle occupants getting distracted from driving-related tasks, so it may be necessary to introduce at new modalities to achieve effective pedestrian-vehicle communication. This research proposes using the lateral deviation of the automated vehicle within its lane as a method to communicate if it is going to yield to the pedestrian. In a crowdsourced experiment, videos containing an approaching automated vehicle were shown to participants. The effect of 1) levels of deviation (no deviation, deviation of 0.4 m, 0.8 m and 1.2 m), 2) direction of deviation (towards pedestrian, away from pedestrian), 3) vehicle behaviour (yielding, not yielding), 4) onset of deviation (onset at a distance of 50 m and 30 m from the pedestrian), and 5) intended vehicle path (vehicle going straight, vehicle taking a turn) were studied. A between-subjects design was used to assign participants (total N = 945; and after filtering, N = 638) randomly to one of 4 groups based on 1) the deviation direction-behaviour mapping (deviation towards pedestrian was ‘yielding’ and away was ‘non-yielding’, deviation towards pedestrian was ‘non-yielding’, away was ‘yielding’) and 2) instructions at the experiment start (not informed of the vehicle deviation, informed of the deviation). Each participant viewed 28 videos, and the task was to press and hold a key as long as it felt safe to cross. The results showed that 1) vehicle deviation to indicate yielding led to a statistically significant improvement in willingness to cross, for all four groups 2) the deviation level was significant when comparing the two extreme values (0.4 m and 1.2 m) for the yielding trials in two of four groups, 3) for one group there was a statistically significant difference in willingness to cross when the vehicle indicates that it intends to take a turn versus when it intends to go straight ahead. It is concluded that lateral deviation of the vehicle, when used to communicate yielding intent. affects the pedestrians’ willingness to cross the road.

Lane change decision-making is an important challenge for automated vehicles, urging the need for high performance algorithms that are able to handle complex traffic situations. Deep reinforcement learning (DRL), a machine learning method based on artificial neural networks, has recently become a popular choice for modelling the lane change decision-making process, outperforming various traditional rule-based models. So far, performance has often been expressed in terms of achieved average speed, absence of collisions or merging success rate. However, no studies have investigated how humans will react to the resulting behavior as potential occupants. This study addresses this research gap by validating a self-developed DRL-based lane changing model (trained using proximal policy optimization) from a technology acceptance perspective through an online crowdsourcing experiment. Participants (N=1085) viewed a random subset of 32 out of 120 videos of an automated vehicle driving on a three-lane highway with varying traffic densities featuring our proposed model or a baseline policy (i.e. a state-of-the-art rule-based model, MOBIL). They were tasked to press a response key if the decision-making was deemed undesirable and subsequently rated the vehicle's behavior along four acceptance constructs (performance expectancy, safety, human-likeness and reliability) on a scale of 1 to 5. Results showed that the proposed model caused a significantly lower amount of disagreements and was rated significantly higher on all four acceptance constructs compared to the baseline policy. Moreover, considerable differences between individual disagreement rates were observed for both models. Our findings offer prospects for the practical application of DRL-based lane change models in a use-case scenario, depending on the user. Further research is necessary to examine whether these observations hold in other (more complex) traffic situations. Additionally, we recommend combining DRL with other modelling techniques that allow for personalization of behavioral parameters, such as imitation learning.

Previous research showed that perceived risk is an important psychological determinant of road user behaviour and accident prevalence. However, little knowledge exists about how objective in-scene features affect a driver’s perceived risk in interactions with pedestrians. This crowdsourcing study tries to fill this research gap. A total of 1082 participants watched 35 out of a total of 86 dashcam videos featuring interactions with pedestrians extracted from the Pedestrian Intention Estimation (PIE) dataset. The videos contained annotations of pedestrian eye contact, crossing behaviour, GPS location, vehicle speed, and yielding rules. The distance between vehicle and pedestrian was manually added, and object counts (detected number of pedestrians, cyclists and vehicles) and respective sizes were added as an index of visual clutter. In each video, participants were asked to press a key on their keyboard and hold it as long as they felt a situation could become risky, and after each video rate perceived risk using a slider and answer whether the pedestrian had made eye contact. Videos in which the participant observed eye contact, increased perceived risk, suggesting that eye contact increases drivers’ vigilance. Videos with more visual clutter, and with higher vehicle speed were also associated with increased perceived risk. However, the causality of the correlation with vehicle speed can be questioned and may be mediated by the environment and whether crossing occurred. Videos in which yielding rules were absent, compared to videos in which they were present, did not affect perceived risk. This study is the first to investigate how pedestrians’ eye contact affects drivers’ perceived risk. The presented results could be useful in safe road design or be used as input for eHMI activation to enhance safety.