Using context to identify difficult driving situations in unstructured environments

Abstract

We present a context-based machine-learning approach for identifying difficult driving situations using sensor data that is readily available in commercial vehicles. The goal of this system is improve vehicle safety by alerting drivers to potentially dangerous situations. The context-based approach is a two-step learning process by first performing unsupervised learning to discover meaningful regularities, or "contexts," in the vehicle data and then performing supervised learning, mapping the current context to a measure of driving difficulty. To validate the benefit of this approach, we collected driving data from a set of experiments involving both on-road and off-road driving tasks in unstructured environments. We demonstrate that context recognition greatly improves the performance of identifying difficult driving situations and show that the driving-difficulty system achieves a human level of performance on cross-validation data. © 2009 Springer.