Learning for Automotive Collision Avoidance and Autonomous Control

Every year 15,000 people are killed on US highways in single vehicle roadway departure automobile crashes, often because of driver drowsiness or inattention. This talk will describe techniques we are developing to prevent these tragedies. Learning is central to this work due to the extreme variability of environmental conditions and individual driver styles. I will describe our recent progress on vision systems that learn to locate the road ahead, and either warning the driver or take control if he is driving inappropriately. I will also describe recent progress in customizing these collision avoidance systems to individual driver characteristics. This customization involves learning a model of typical driver steering behavior, and only taking action when the current situation falls outside of a particular driver's normal operating envelope. This technique shows promise for reducing false alarms, while maintaining system vigilance. I will present results from experiments with these warning and control techniques on a wide variety roads at speeds of up to 70 mph.