dc.description.abstract | In recent years, convolutional neural networks (CNNs) have been applied to several
autonomous driving tasks in an end-to-end manner with promising results. As these
kinds of systems are given the freedom to self-optimise to maximise their overall
performance, they are believed to eventually reach better performance than more
fragmented solutions, where intermediate, human-selected criteria are optimised
instead. End-to-end systems drastically reduce the amount of manual labour needed
to annotate datasets.
This thesis explores end-to-end learning for autonomous driving using a SPURV
Research robot. A CNN architecture trained on self-collected data is used to predict
steering angles for three respective tasks: Indoor lane following using the SPURV
robot, lane following inside the Udacity Simulator, and finally, a task in which a
U-turn manoeuvre, which requires a notion of history, is performed. The goal is to
check if this can be done with only feed-forward neural networks. The models for
the first and last tasks are tested in real-life on the SPURV robot.
The indoor lane following was successful, and the SPURV robot followed lanes
correctly and stably. The marker turning model showed less promising results: The
SPURV was able to complete the task occasionally but with several failures. The
SPURV vehicle was found to be a useful framework for data collection and testing of
methods for autonomous driving and revealed that the quantitative metrics used in
this project did not accurately reflect real-life performance. | |