Real-time driver model using an Unscented-Kalman Filter for driver characterization

This Ph.D. thesis presents a framework for characterizing drivers by estimating a set of parameters using an Unscented Kalman Filter (UKF) in conjunction with a dynamic model. While existing academic work primarily focuses on studying motorway and car-following behaviour, this research addresses the need for understanding pure driving on country roads, which constitutes the majority of everyday driving scenarios. Two pure driving studies (henceforth called naturalistic driving study) were conducted, providing valuable real-world data and overcoming limitations associated with simulator studies.

The UKF was chosen as the state estimation algorithm due to its familiarity, real-time estimation capabilities, and improvements in computational efficiency. The thesis investigates the applications and outcomes of linear, Extended, and Unscented Kalman filters using a quarter-vehicle model as an example system, with body mass as a parameter to be identified. The comparative analysis reveals that the linear Kalman filter is susceptible to rigid body drift when applied to nonlinear systems, while the UKF is preferred for this application over the EKF as it eliminates the need for computing system Jacobians, but both are effective solutions. The UKF demonstrates robust performance for nonlinear systems, and its computational efficiency is further improved in subsequent stages of the research.

The thesis proceeds by presenting the vehicle, lateral, and longitudinal controllers. The vehicle model employed is a bicycle model, which adequately represents normal driving situations where vehicle dynamics are not highly influential. The lateral controller parameterizes driver behaviour through lookahead time and understeer gradient estimate, while the longitudinal controller is specifically developed to emulate natural open-loop manoeuvres performed by drivers when negotiating corners and speed limits. It parameterizes driver behaviour in terms of specific acceleration event timings and magnitudes. The model accurately captures the timing and impulse of acceleration throughout the manoeuvre, and one parameter is transformed into an estimate of a driver's lateral acceleration tolerance during cornering, enhancing the model's accuracy. Straight-line cruising behaviour in speed-limited or derestricted sections is handled through an auto-regressive moving average that tracks the driver's recent preferred speed on straight sections.

The subsequent chapter demonstrates the implementation of the UKF with the lateral controller and its capability to adapt parameters accurately in real-time using only vehicle steering angle information. The method is validated using simulated signals and showcases the filter's ability to compensate for vehicle position errors through lookahead time adjustment. This feature of the algorithm aligns the sensor and model signals, reducing discrepancies caused by GPS position inaccuracies. Importantly, this highlights the algorithm's capacity not only to identify parameters accurately but also to mitigate errors arising from imperfect sensor measurements. The chapter also establishes that the UKF can operate with an offline vehicle model, resulting in a smaller state vector, reduced computational load, and seamless integration of both longitudinal and lateral controllers with the UKF.

The integration of the UKF with the longitudinal controller is presented in the subsequent chapter. The filter's real-time operation is demonstrated using readily-available acceleration and steering angle signals, capturing variations in driver behaviour caused by changes in mood or specific journey conditions. Notably, the thesis introduces the estimation of time-based parameters, which serve as important indicators of driver behaviour when correlated with factors such as energy wastefulness and safety. The thesis explores the relationship between specific driver measures, such as time spent in acceleration bands, steering variation, and jerk, and their correlations with model parameters. This part of the work has also been published as a Journal Paper in IMechE Part D Journal of Automobile Engineering. Furthermore, this chapter also explores the effects of increased or reduced cognitive load introduced through distractions or classical music as a means of making the driver more relaxed. The findings have been presented at the AVEC 2022 conference and published as a conference paper.

The next chapter explores specific metrics such as lateral and longitudinal accelerations, percentage of time spent above specific acceleration and jerk thresholds in addition to further steering and energy expenditure metrics. Pearson linear correlation coefficients are calculated for all possible combinations, revealing significant correlations between metrics such as overall average longitudinal velocity on straight sections, post-corner acceleration magnitude, lateral acceleration tolerance when negotiating corners, and energy expenditure per kilometer. These correlations support the fits to energy and safety metrics discussed in the previous chapter, showcasing the ability of a subset of model parameters to accurately predict these metrics and effectively span the behaviour range of drivers.

In a second driving study focused on driver aggressiveness, specific predictors are developed based on the speed with which drivers negotiate road features with higher accident risks. Additionally, driver reaction times are tested, accompanied by a survey to provide a means to correlate driving behaviour with selfevaluation. The drivers chosen for this study were classified based on their cumulative rank order of parameters highly correlated with energy expenditure. The study reveals a wide range of responses, with extreme behaviours exhibited by specific subsets of drivers, aligning with the initial classification. These findings motivated a revisit of the model presented in Chapter 7 to identify predictors of aggressiveness from the driving study results. The results confirm the predictive capability of the model by validating it against data from the second driving study, accurately predicting drivers' actual aggressiveness levels. The survey and reaction time tests uncover intriguing results, including improvements in reaction times among female participants and self-perceptions of faster reaction times that diverge from test results.