(A) platform-independent hybrid control architecture for robust autonomous driving

Abstract

In this thesis, a platform-independent hybrid control architecture that generates control commands for lateral and longitudinal autonomous driving is proposed. Based on the deep neural network for the basic lane keeping autonomous driving, the End-to-End learning method which learns the nonlinear relation from the driving environment information to the final control command is applied to make robust driving in various driving environments. In addition, ICP matching algorithm is applied between driven trajectory using real time perception and global plan from start to destination, so it can compensate localization errors occurring during driving. Also, in order to solve the dependence on the platform which is a limitation of End-to-End learning approach, image transformation and vehicle modeling are performed. The performance of proposed architecture is verified through experiments in a real road environment.