Detection of Data Matrix Encoded Landmarks in Unstructured Environments using Deep Learning

Visual based navigation in industrial environments has been a challenge because of the complex and unstructured nature of the surroundings that Automated Guided Vehicles (AGV) and autonomous robots have to traverse in a continuous routine in many production facilities. Since natural landmarks are still a huge challenge due to the highly dynamic configuration of the environment and the varying nature of these landmarks, simple and low cost artificial landmarks appear as a potential solution. The one exploited in this work uses simple sheets of paper with Data Matrix encoded markers spread in the environment trying to create a constellation of landmarks in such a way that several of them are always visible to a set of cameras on board the robot or AGV. All codes are unique, which makes robot continuous localization a much simpler challenge whenever at least two or three landmarks are visible. When using traditional vision techniques in large images of the scene, one of the most demanding parts is to detect the landmarks to further process them. For that purpose, this paper proposes a technique based on deep learning that efficiently detects these special landmarks in images. A dedicated dataset was created and a Faster R-CNN architecture was adapted and trained for that purpose. The results show that almost all markers was detected in the images with a processing speed larger more than one order of magnitude than traditional techniques, including in demanding situations of poorer illumination or partial occlusions.