“Fitting Data”: A Case Study on Effective Driver Distraction State Classification

Abstract

he goal of this project was to investigate how to make driver distraction state classi cation more e cient by applying selected machine learning techniques to existing datasets. e data set used in this project included both overt driver behavior measures (e.g., lane keeping and headway measures) and indices of internal cognitive processes (e.g., driver situation awareness responses) collected under four distraction conditions, including no-distraction, visual-manual distraction only, cognitive distraction only, and dual distraction conditions. e baseline classi cation method that we employed was a support vector machine (SVM) to rst identify driver states of visual-manual distraction and then to identify any cognitive-related distraction among the visual-manual distraction cases and other non-visual manual distraction cases. e new aspect of this research is optimization of the classi cation e ort, which involved cardinality constraints on 16 overt driver behavior measures. A spline transformation was also implemented to achieve better classi cation performance. In addition to testing our optimization approach with the SVM, we also explored logistic regression. Results revealed the splinetransformed variables to produce a good "out-of-sample" performance for both the SVM and logistic regression. Beyond this, the cardinality constraints were important for selection of the most in uential variables in driver state classi cation accuracy and preventing data over tting. Regarding the objective of eciency in distraction classi cation, with only two input variables our optimized approach achieved state classi cation accuracies similar to accuracies achieved with "brute-force" application of SVM with all 16 overt driver behavior measures as inputs. Interestingly, with splined-transformed variables, reducing the number of input variables to 2 only led to a marginal decrease in classi cation accuracy (75.38% to 74.16%). e optimization methods explored in this paper could be applied to other in-vehicle real-time data to reduce computational demands in using machine learning methods for driver state classi cation.