ArUco Marker Detection under Occlusion Using Convolutional Neural Network

Abstract

Camera pose estimation is a significant warranty during Unmanned Aerial Vehicle (UAV) autonomous landing process. Fiducial marker system is a popular method to offer relatively precise pose information. However, square-based markers are unreliable under occlusion condition, especially when their corners are covered by unexpected disturbances. This study proposes a novel method to detect fiducial markers using neural network. The method is developed based on Convolutional Neural Network (CNN) and achieves outstanding results under various occlusion conditions, including different cover shapes and ratios. YOLOv3, along with its improved version YOLOv3-spp and its lightweight version YOLOv3-tiny, are applied as the marker detector. Compared to the traditional ArUco fiducial marker system, CNN architectures are more robust and stable in extreme environment. Performance of three different CNN models is quantified as marker detection rate. This work validates the feasibility of square-based fiducial marker localization employing CNN architecture, and reveals the potential of deep learning method in the field of fiducial marker detection and recognition.