Evaluation and comparison of parametric and non-parametric methods for driving behavior analysis

Car crashes are a scourge to society world-wide and a significant public health issue. Many studies have been conducted using different statistical techniques to understand the causal factors for these crashes, and to minimize injuries and deaths. This thesis aims to identify and analyze the techniques used in previous studies, propose better techniques and demonstrate that these techniques can be applied to build better models which yield us better prediction results. Chapter 2. In this chapter, data from a naturalistic driving study was used to examine drivers’ seat belt buckling time and behavior based on the driver’s demographics and drive conditions. Parametric methods were applied to these data. Logistic regression analysis was performed on this data, and appropriate model was built to represent the driver’s seatbelt buckling behaviors. Linear regression technique was used to build models that can predict the driver’s seatbelt buckling time during driving. Thus, with the help of these models, driver’s seat belt buckling behavior and buckling time can be predicted based on the driver’s demographics and drive conditions. The results from these regression techniques gave us better insights of the data used in the analysis. For example, Age and Sex were identified as representative contributors in buckling time analysis and the mean buckle time of 14 seconds was determined from the analysis. Moreover, the results demonstrated that as the driver’s age increases by one year the buckle time increases by 0.044 seconds and male drivers tend to buckle faster by 4.072 seconds than female drivers. Similarly, such interesting results were obtained from buckling behavior analysis which were discussed in this thesis. With the models developed from these regression techniques seat belt reminder systems can be customized which can reduce the perception of nuisance and protects the drivers from higher speed unbuckled crashes. Chapter 3. In this chapter, a different data set was used, this time from the National Advanced Driving Simulator from a set of ‘legacy data’ to study the relationship between driving performance measures (DPMs) and driving conditions. DPMs were collected under various controlled driving conditions to demonstrate different driving behaviors. Given the limited number of controlled driving conditions in experimental data, additional methods to model and predict the DPMs under unobserved driving conditions were modeled. Although interactions among different DPMs are studied and are widely reported in existing literature, these interactions were not fully considered in DPM modeling. The techniques used in previous literature considered the analysis of a single DPM at a time (each DPM is modeled individually), even though modeling multiple DPMs together by considering the interactions among DPMs is beneficial. To expand previous works, a novel DPM modeling and prediction method i.e., multi-output convolutional Gaussian process (MCGP) which can incorporate the interactions among different DPMs was proposed in this chapter. This method features the modeling flexibility for different DPMs and the interpretable modeling structure for integrating the DPM interactions. Unlike in chapter 2, where just the application of parametric methods was demonstrated, in this chapter parametric and non-parametric methods were compared, and their performances were evaluated. The proposed method is compared with three benchmark methods- Generalized linear model, Specific linear model, and Univariate Gaussian process on the DPM data set, and the results demonstrate the superiority of the MCGP method. Discussions and interpretations of the results are also provided. These studies on the seat belt buckling data and the DPM data help us to identify the appropriate method based on the data being analyzed. The advantages and disadvantages of the parametric and non-parametric methods are discussed in each application scenario to demonstrate the applicability of each method. Moreover, insightful results are discovered after analyzing the seat belt buckling data and the DPM data. For example, age and sex are identified as the important independent variables when modeling the seat belt buckling time. The results from DPM data analysis were used to demonstrate the superiority of modeling related DPMs together, which presents an interpretable characterization of the interactions among different DPMs. The proposed methods in this thesis are also applicable in many other driving behaviors, e.g., driving anomaly detection and prediction.