Title: Visual Attention for High Speed Driving

Committee:

Dr. James M. Rehg, CoC, Chair

Dr. Evangelos Theorodou, AE

Dr. Byron Boots, CoC

Dr. Dhruv Batra, CoC

Dr. Dieter Fox, IC

Abstract:

Coupling of control and perception is an especially difficult problem. We chose to study this problem in the context of aggressive driving, and propose using a learned deep neural network attention mechanism to compare our learned gaze strategy to a human attentional strategy. We show that a  convolutional neural network can directly learn a mapping from input images to top-down cost map. This cost map can be used by a model predictive control algorithm to drive aggressively. We further show the ability to learn an end to end trained gaze neural network gaze strategy that allows both high performance and better generalization at our task of high speed driving. We compare this gaze with that of a human driver performing the same task.  Using these methods, we demonstrate repeatable, aggressive driving at the limits of handling on a physical robot.