Extraction of descriptive driving patterns from driving data using unsupervised algorithms

Abstract

Understanding drivers’ behavioral characteristics is critical for the design of decision-making modules in autonomous vehicles (AVs) and advanced driver assistance systems (ADASs). Current relevant studies are mainly based on supervised learning methods which involve extensive human efforts in model development. This paper proposed a framework for automatic descriptive driving pattern extraction from driving sequence data using unsupervised algorithms. Based on the Bayesian multivariate linear regression model, two unsupervised algorithms were utilized to segment driving sequences into fragments. Three extended latent Dirichlet allocation models were applied to cluster the fragments into multiple descriptive driving patterns. The collected driving data from a naturalistic driving experiment was applied to examine the effectiveness of our proposed framework. Results show that the unsupervised segmentation algorithms could help effectively detect the switch characteristics between two continuous driving maneuvers along time, and the clustered patterns could effectively describe the characteristics of each driving maneuver. The proposed unsupervised framework provides an effective and efficient data mining solution to facilitating deep and comprehensive understanding on drivers’ behavioral characteristics, which will benefit the development of AVs and ADASs.