Machine Learning for Lane Positioning in Autonomous Vehicles

Abstract

The lateral position of a car in its current lane affects the perceived comfort and safety of passengers. The problem of selecting a comfortable lateral position for an autonomous car in traffic is not trivial. This thesis addresses this issue by breaking it down into two related problems to be solved, using Feed-Forward Neural Networks and Random Forests. The first problem, Barrier Detection (BD), is a perception problem. It involves identifying relevant variables for detecting barriers on the road side and distinguishing barriers from instances resulting in similar sensor stimuli such as traffic signs. The second problem, Lane Positioning (LP), is a Decision & Control (D&C) problem. The task is to select a suitable lateral position for the host vehicle within the current lane, given the immediate surroundings. Furthermore, strict timing and memory requirements are in place. This necessitates pruning and preprocessing of high-dimensional raw data before the data is forwarded, and places constraints on the implementations. Both problems were approached with supervised learning methods, and the results for BD were promising with a test set accuracy of 80% using neural networks. The neural networks for LP failed, with the best neural network attaining only 60% accuracy. By using a different classifier type, Random Forest (RF), an out-of-bag1 error of 4% was achieved for the LP problem. However, sufficient performance was only reached for extreme lateral offsets and these classifiers grossly violated the memory constraints. Random lateral offsets in the data appeared to overshadow the offsets taken in response to significant scenarios. This could explain why the classifiers failed to identify trends for smaller offsets.