Pattern matching as the nucleus for either autonomous driving or driver assistance systems

Concerns autonomous vehicle driving by pattern matching combined with reinforcement learning. In specific, this research focuses on the requirement to steer an autonomous car along a curvy and hilly road course with no intersections and no other vehicle or obstacle but with the strict requirement to self-improve driving behaviour. A camera is used to build quickly an abstract complete description (ACSD) of vehicle's current situation. This combines traditional edge finding operators with a new technique of Bayes prediction for each part of the video image. Those ACSD's are being stored together with the steering commands issued at that time and serve as the pattern database of possible driving behaviour which are being retrieved using an approximate nearest neighbour pattern matching algorithm with a O(n log m) characteristic compared to O(n/spl middot/m) for the conventional nearest neighbour calculation. In addition to this, any feedback on the quality or appropriateness of the driving behaviour has to be self-created (e.g. time measurement for a whole road section) and is therefore delayed and unspecific in relation to single issued steering commands. Consequently, a machine learning algorithm coping with those conditions is being implemented based on Reinforcement Learning.