Authors:
Sandesh Hiremath
;
Praveen Gummadi
;
Argtim Tika
;
Petrit Rama
and
Naim Bajcinca
Affiliation:
Department of Mechanical and Process Engineering, Rheinland-Pfälzische Technische Universität Kaiserslautern-Landau, Gottileb-Daimler-Straße 42, 67663 Kaiserslautern, Germany
Keyword(s):
Autonomous Driving, MPC, Learning Based Control, End-to-End Control, Lane Detection.
Abstract:
This work proposes a learning-based controller for an autonomous vehicle to follow lanes on highways and motorways. The controller is designed as an interpretable deep neural network (DNN) that takes as input only a single image from the front-facing camera of an autonomous vehicle. To this end, we first implement an image-based model predictive controller (MPC) using a DNN, which takes as input 2D coordinates of the reference path made available as image pixels coordinates. Consequently, the DNN based controller can be seamlessly integrated with the perception and planner network to finally yield an end-to-end interpretable learning-based controller. Here, all of the controller components, namely- perception, planner, state estimation, and control synthesizer, are differentiable and thus capable of active and event-triggered adaptive training of the relevant components. The implemented network is tested in the CARLA simulation framework and then deployed in a real vehicle to finally
demonstrate and validate its performance.
(More)