Developing a model-based lane change decision aid system by integrating driver uncertainty

Abstract

As a complex and highly dynamic driving task, the lane change maneuver is considered as one of the most dangerous driving maneuvers. It requires drivers to integrate highly dynamic information from different sources to make a safe decision and to perform the maneuver safely in a timely manner. Driver uncertainty about the current situation can substantially prolong this decision-making process, potentially leading to dangerous lane change maneuvers. Regarding traffic safety, it is therefore essential to consider driver uncertainty while developing lane change assistance systems. In addition, the information that existing lane change assistance systems provide to the drivers is not adaptive to drivers’ uncertainty states during decision-making in a current situation. Without taking driver uncertainty into account, this non-adaptive assistance can be either obtrusive with too many unnecessary advice or too unobtrusive without providing needed information on time. For instance, when drivers are quite certain about the situation and systems still provide information actively, the interaction with such systems can be annoying and even disturbing. When drivers are uncertain during the decision-making process and systems provide no helpful information on time, the interaction with the systems will be not beneficial. Such non-adaptive aids may have an impact on drivers’ mental models of the system functionality and further decrease driver trust in assistance systems. As a consequence, it can result in the disuse of such systems. Hence, it is also important to build driver trust in automation with the consideration of driver uncertainty while developing lane change assistance systems in addition to providing traffic safety. The “emancipation” theory of trust states that humans tend to trust in their partners when they are uncertain in social interaction. Inspired by this, it is assumed that in the context of humanmachine interaction, drivers tend to build trust in systems and to avoid the occurrence of system disuse, when systems consider driver uncertainty and help to shorten the decision-making process and the decision times accordingly. Based on this assumption, this dissertation aims to develop a Model-Based Lane Change Decision Aid System (MBLCDAS) integrating driver uncertainty during decision-making in its interaction strategy for lane change maneuvers. The MBLCDAS adapts its information behavior to driver uncertainty: Information will be given in an active way when drivers are uncertain during decision-making, and information will be presented unobtrusively when drivers are certain. This adaptive assistance can help build appropriate trust in the assistance system and also provide traffic safety. To develop the MBLCDAS, first driver uncertainty during decision-making has been studied in a driving simulator for specific lane change scenarios on two-lane motorways. The distance gap, closing speed and time to collision (TTC) between the subject vehicle and approaching vehicle are varied and found to significantly influence driver uncertainty. Reaction times, subjective uncertainty scores and the action proportion for lane change decisions are used to measure driver uncertainty. With a frequency of 60 Hz in the driving simulator, information of current traffic situations concerning distance gap, closing speed and the resulting TTC between the subject vehicle, the lead vehicle, and the approaching vehicle is collected and then used to develop a probabilistic model of driver uncertainty. Regarding the attributes for the model of driver uncertainty, the TTC between the subject and the lead vehicle as well as the approaching vehicle, the distance gap between the subject and the lead vehicle as well as the approaching vehicle are selected based on the mutual information. For the outputs of the model, the collected subjective uncertainty scores from the driving simulator experiment are mapped onto the conceptualized two uncertainty states of the MBLCDAS based on the Kullback-Leibler Divergence. Regarding the structure of the model, Tree-Augmented Naive (TAN) Bayesian classifier is selected among the candidate models (full Bayesian classifier, naive Bayesian classifier, TAN classifier) with the highest Bayesian Information Criterion (BIC) score. After learning the structure and parameters of a TAN Bayesian classifier, the conditional probability of the driver uncertainty in a given lane change situation can be inferred. The developed model of driver uncertainty is then evaluated with test data, showing an average accuracy of approximate 0.78. Based on the selected decision thresholds, the inferred probability of driver uncertainty can be classified into the driver’s uncertainty state as either “certain” or “uncertain”. In addition, with the help of the action threshold between the subject and the approaching rear vehicle, the recommendation for lane change decisions can be classified as either “decision for changing the lane” or “no lane changes decision” by a safety analysis. The classified driver’s uncertainty state and the classified recommendation for lane change decisions together trigger the Human Machine Interface (HMI) via symbols representing different levels of criticality and driver uncertainty. Emotional faces consisting of the dimensions of colors and emotional expressions are chosen as symbols: Colors describe criticality as well as driver uncertainty, while emotional expressions provide recommendations for lane change decisions. After implementing the model of driver uncertainty and the corresponding HMI in the driving simulator, the developed MBLCDAS has been then evaluated with 20 participants. In the evaluation study, the MBLCDAS has been compared with other reference systems with respect to the reduction of reaction times and the building of trust. The results show that all systems including the MBLCDAS are able to reduce reaction times in comparison to the driving without any assistance. In addition, trust has been built in the MBLCDAS. Compared to other reference systems without considering driver uncertainty, the MBLCDAS is reported to be most accepted and trusted by participants.