Authors:
Masoud Hoveidar-Sefid
and
Michael Jenkin
Affiliation:
Electrical Engineering and Computer Science Department and York Centre for Field Robotics, Lassonde School of Engineering, York University, Toronto and Canada
Keyword(s):
Autonomous Navigation, Trail Following, Path Finding, Deep Neural Networks.
Related
Ontology
Subjects/Areas/Topics:
Autonomous Agents
;
Image Processing
;
Informatics in Control, Automation and Robotics
;
Intelligent Control Systems and Optimization
;
Neural Networks Based Control Systems
;
Robotics and Automation
Abstract:
Trails are unstructured and typically lack standard markers that characterize roadways; nevertheless, trails can provide an effective set of pathways for off-road navigation. Here we approach the problem of trail following by identifying the deviation of the robot from the heading angle of the trail through the refinement of a pretrained Inception-V3 (Szegedy et al., 2016a) Convolutional Neural Network (CNN) trained on the ImageNet dataset (Deng et al., 2009). A differential system is developed that uses a pair of cameras each providing input to its own CNN directed to the left and the right that estimate the deviation of the robot with respect to the trail direction. The resulting networks have been successfully tested on over 1 km of different trail types (asphalt, concrete, dirt and gravel).