Cross Validation for CNN based Affordance Learning and Control for Autonomous Driving

Abstract

Autonomous driving has attracted a significant amount of research efforts over the last few decades owing to the exponential growth of computational power and reduced cost of sensors. As a safety-sensitive task, autonomous driving needs a detailed level of scene understanding of decision making, planning, and control. This paper investigates the Convolutional Neural Network (CNN) based methods for affordance learning in driving scene understanding. Various perception models are built and evaluated for driving scene affordance learning in both the virtual environment and real sampled data. We also propose a conditional control model that maps the extracted coarse set of driving affordances to low-level control condition on the given driving priors. The performance, merits of the CNN based perception models, and the control model are analyzed and cross-validated on both virtual and real data.