Presentation

· Contributors · Organizations · Search Program · Flagged · Happening Now

Applying Information Theory Divergence to Unsupervised Clustering of Driving Gaze Distribution

DescriptionTracking driver’s gaze is an integral part of driving monitoring systems, designed to assess the driver’s state. In this work, we report a novel method that creates a ‘vocabulary’ of driver’s gaze patterns during naturalistic driving, across a variety of driving environments. We use an information-theory based unsupervised clustering method to create a vocabulary that contains mutual information between properties of the driving scenarios (e.g. approaching and crossing junctions) and the gaze clusters. Our clustering begins with a dual-staged process of a change of representation. It transforms segments from raw gaze samples into a probabilistic Gaussian Mixture Model (GMM). By using GMM, we were able to represent a gaze sequence of several seconds as a stationary distribution. These GMMs were then grouped in the second stage using information-theory divergence into clusters of segments. The cluster IDs are our ‘vocabulary’. We evaluate our model on a dataset of 15 hours of naturalistic driving, comprising urban, suburban, and highways. The evaluation demonstrates that our clusters preserve relevant driving-related information, and the vocabulary contains meaningful states of driving behavior.

Contributors

Ron M. Hecht

General Motors