Study on visual machine-learning on the omnidirectional transporting robot

Abstract

We present a computer vision solution integrated to an omnidirectional transporting robot to perform the position tracking of multiple trays moving on its planar acrylic plate surface. The trays were designed to carry lightweight materials on top of their surface so that the mechanism could be implemented as an automated transporting system for applications that require the displacement of products and/or materials in any given space. One hurdle faced by the visual system for suitable detection was the partial occlusion of the image of a tray when placing arbitrary objects on its surface. Our strategy to overcome this challenge consisted on the implementation of machine learning algorithms, such as Support Vector Machines (SVM), using datasets of images containing trays with different occlusion patterns for fast object detection through rigorous training. The results of experimental tests validate the implementation of our proposal as a reliable approach for the object tracking of multiple trays on the robotic device, even under partial occlusion. We also studied the accuracy of the position measurements performed by our visual system with respect to the position measurements taken by the OPTITRACK motion capture system and evaluated the processing time per frame required by the software implementation.