Authors:
Qazi Jan
;
Jan Kleen
and
Karsten Berns
Affiliation:
Robotic Research Lab, Department of Computer Science, Reinland Pfälzische-Technische Universität, Erwin-Schrödinger-Straße 52, 67663 Kaiserslautern, Germany
Keyword(s):
Autonomous Driving, Deep Driving, Pedestrian Zones, Simulation, Neural Networks, Directional Inputs.
Abstract:
Deep Neural Networks are being used in different applications to solve complex tasks with high precision. One application, also the focus of this paper, is end-to-end driving. Generally, in an end-to-end approach, a neural network learns to directly feed values to actuators based on sensor inputs. This paper uses an End-to-end approach with images and additional direction inputs:left, right and straight for imposing a certain direction at unstructured and arbitrary intersections of pedestrian zones. Expecting high precision for predicted steering in pedestrian zones could be uncertain due to the atypical structures of intersections. Findings for increased accuracy are done using direction inputs with three variants of two approaches: Single and parallel model. Depth information was included to overcome shadow problems from RGB in simulation, but it resulted in worsening the drive, and hence removed in further experiments. The experiments are performed in simulation to verify the util
ity of the proposed approaches and narrow down the best models for actual hardware. From the experiments, it is seen that parallel model with front images have performed best. The model drove well along the paths and followed the given input direction from the user at the crossings. To maintain the length of this paper, only results for parallel structures are discussed.
(More)