Autonomous Vehicle End-to-End Reinforcement Learning Model and the Effects of Image Segmentation on Model Quality

Abstract

Autonomous driving has the potential not only to transform people's lives, but also save them. Fully understanding state of the art autonomous driving architectures, however, requires a wide breadth of knowledge on available sensors, perception, image segmentation, localization, path planning, neural networks, convolution, and more. This thesis proposes a simple end-to-end architecture that has promising behavioral results. Two novel techniques are also introduced: a new exploration algorithm that seeks to produce more robust training behaviors over simple linear decay models, and a new data splitting technique that splits a layer into multiple semantically meaningful layers in an attempt to improve feature recognition by a convolutional neural network. A series of end-to-end models are trained with access to either a ground truth semantic segmentation perceptor or an image camera perceptor with a semantic segmentation predictor model. Models are evaluated and results are compared to see which approach is superior. The perceptor configuration on the trained models is switched and evaluation is run again to see how the it reacts to a change in perceptor quality. This thesis hypothesizes that models should be trained on ground truth semantic segmentation data, even if the trained model will ultimately be evaluated with a semantic segmenter model, as the model quality should prove superior and training time can be reduced substantially.