Driver-Adaptive Lane Departure Warning Systems

Each year, there are thousands of car accidents in the U.S. These accidents claim many lives, and cost billions of dollars. There are many different types of accidents, including rear end collisions, side swipes, head on collisions, collisions with static obstacles, accidents while merging or changing lanes, and driving off the road. Mandatory seat belt usage, air bags, lower speed limits, rumble strips, and stricter vehicle safety requirements have all helped to reduce the number of accidents and fatalities. However, it is now possible to do more, by using intelligent driver assistant systems. In this thesis, I concentrate on a particular type of accident, known as Run-Off-Road (ROR). An ROR crash occurs when a single vehicle departs the road, due to either driver inattention, drowsiness, or other incapacitation, and then impacts something, such as a tree or a house. Previous work in preventing ROR accidents mostly makes use of Lane Departure Warning Systems, which are usually vision-based lane trackers. These systems predict when the driver is in danger of departing the road, and trigger an alarm to warn the driver. To warn drivers early enough so they have time to react means that often, alarms are generated in situations where there is no real danger of a crash. These alarms are called nuisance alarms. My goal is to reduce the number of nuisance alarms, while maintaining adequate time for the driver to respond to a truly dangerous situation. Using real world driving data, I show that achieving this goal requires more intelligent modelling of the driver’s behavior than most current systems are capable of. This modelling comes in two forms: A novel “alarm decision model,” which takes into account road geometry and past driver behavior, and a training algorithm which tunes certain model parameters to an individual driver. This new model reduces nuisance alarms, while maintaining adequate warning time, for “loose” drivers who weave excessively. The improvement for “tighter” drivers is less, as current warning systems already do a good job on them. I analyze the reason for improved warning system performance using a memory based learning framework, and show why “loose” drivers are helped more than “tight” drivers.