Multi-LED Classification as Pretext For Robot Heading Estimation

Our chosen model is a Vision-based relative pose estimation of peer robots is a Fully Convolutional Network (FCN) [8] that, given an image, key capability for many multi-robot applications [1]. While predicts two-dimensional maps: one for the state of each some approaches focused on handcrafted algorithms, based LED, one representing the model's belief about the robot on circles printed on paper sheets [2] or exploiting the image-space position, and one about its heading. We take LEDs geometry [3], these systems lacked adaptability to full advantage of the inductive bias of FCNs, in which the different environmental conditions. Recent solutions train value of a pixel in the output map depends on a limited deep neural networks (DNNs) on large amounts of data.