AbstractThis thesis investigates the influence of different driving style on fuel consumption through the development of data-driven personalised driver models. A systematic study was conducted which consists of five main components, ranging from literature review to driver modelling and evaluation.
Two literature reviews are included in this thesis. The first review focuses on exploring the potential connections between driving style and fuel consumption, which confirms the motivation of this research. Meanwhile, the second review evaluates the existing studies on driver modelling that addressed human factors. The adopted modelling approach and feature parameters in this study are justified through the comparisons of existing studies in this second review.
Afterwards, a novel data-driven personalised driver modelling approach is presented, together with driving data collection, driving style classification and simulation evaluation. The major contributions of this thesis lay in six aspects. Firstly, a novel visual sensing algorithm is created to detect the potential leading vehicle and estimate the headway distance from the recorded dashcam footage. Secondly, a unique sensor fusion approach based on Kalman filter is developed to fuse the sensory data from radar and dashcam, and hence generate an optimised headway distance measurement. Thirdly, a novel driving style classification approach based on Support Vector Clustering is developed to differentiate the driving style variations. Fourthly, two fuzzy logic calibration approaches are adopted to infer the driver models from the collected data. Fifthly, the established driver models are incorporated into procedures anchored to the standard World-wide harmonized Light duty Test Cycle, to facilitate the investigations on the correlations between driving style and fuel consumption. Finally, a general process of developing personalised driver models is proposed to benefit relevant studies.
The conclusions reached are: (i) the proposed visual sensing algorithm can effectively estimate the headway distance from the recorded footages. (ii) The developed sensor fusion approach can compensate the drawback of each individual sensor with a strong robustness. (iii) The proposed driving style classification approach can differentiate driving style variations with a validated performance through the comparison with the jerk profiles of each participant. (iv) Both calibrated fuzzy driver models outperform the traditional PID-based model in characterizing human driving style. (v) In the proposed simulation scenario, the fuel reductions between the established aggressive and defensive driver models can be approximately 6% and 8% for each calibration approach.
|Date of Award||19 Jun 2019|
|Supervisor||Chris Brace (Supervisor) & Pejman Iravani (Supervisor)|