“Fitting” data: A Case Study of Effective Driver Distraction State Classification 2019-01-0875
The goal of this project was investigate how to make driver distraction classification more efficient by applying selected machine learning techniques to existing data sets. The data set selected for this project included both overt driver behavior measures (e.g., lane keeping and headway measures) and internal process indices (e.g., driver situation awareness responses) collected during four distraction conditions, including no-distraction, visual-manual distraction only, cognitive distraction only and dual distraction conditions. The baseline classification method that we employed was a support vector machine (SVM) to first identify driver states of visual-manual distraction and then to identify the cognitive-related distraction states among the visual-manual distraction cases and other non-visual manual distraction cases. The classification optimization effort, which is the new aspect of this line of research, involved cardinality constraints to 16 overt driver behavior measures. A spline transformation was implemented to achieve better classification performance. In addition to testing our optimization approach with the support vector machine, we also explored logistic regression. Results revealed the spline-transformed 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 influential variables in driver state classification accuracy and preventing data overfitting. Regarding the need for data reduction, with only two input variables, our optimized approach achieved driver state classifications accuracies similar to the level of accuracy achieved with a “brute-force” application of SVM with all 16 overt driver behavior measures as inputs. Interestingly, increasing the number of selected variables from 2 to 3 only led to a marginal improvement in classification accuracy (74.16% vs. 75.38%). The optimization methods explored in this paper could be applied to other in-vehicle real-time data using machine learning methods in order to reduce computational demands.
Yu Zhang, David Kaber, Stan Uryasev, Alexey Zrazhevsky
DENSO International America Inc., University of Florida, American Optimal Decisions, Inc.