End-to-End Steering Angle Prediction for Autonomous Vehicles
Abstract
In recent years, convolutional neural networks (CNNs) have been applied to severalautonomous driving tasks in an end-to-end manner with promising results. As thesekinds of systems are given the freedom to self-optimise to maximise their overallperformance, they are believed to eventually reach better performance than morefragmented solutions, where intermediate, human-selected criteria are optimisedinstead. End-to-end systems drastically reduce the amount of manual labour neededto annotate datasets.
This thesis explores end-to-end learning for autonomous driving using a SPURVResearch robot. A CNN architecture trained on self-collected data is used to predictsteering angles for three respective tasks: Indoor lane following using the SPURVrobot, lane following inside the Udacity Simulator, and finally, a task in which aU-turn manoeuvre, which requires a notion of history, is performed. The goal is tocheck if this can be done with only feed-forward neural networks. The models forthe first and last tasks are tested in real-life on the SPURV robot.
The indoor lane following was successful, and the SPURV robot followed lanescorrectly and stably. The marker turning model showed less promising results: TheSPURV was able to complete the task occasionally but with several failures. TheSPURV vehicle was found to be a useful framework for data collection and testing ofmethods for autonomous driving and revealed that the quantitative metrics used inthis project did not accurately reflect real-life performance.