Visual localization is to estimate the position and rotation of a camera in three-
dimensional space from images, which has wide and important applications in virtual
reality (VR), augmented reality(AR), robot navigation, and unmanned driving. How-
ever, under the conditions of weak textures, motion blur and illumination changes, the
accuracy and robustness of visual localization are not high. Visual localization with
markers is a good solving scheme. Circular markers are common in daily life and have
high discriminative feature. This paper focuses on visual ellipse region detection and its
application on visual localization from circle markers. The contributions of the paper
are as follows:
1. An image ellipse region detector based on a rotating bounding box is proposed,
named EllipseNet. Generally, projections of circular markers in images are elliptical
shapes. The detector can extract the region of the ellipse marker and predict the angle
of the ellipse in any direction. Experiments demonstrate that accuracy of the detector is
higher than those with upright bounding boxes.
2. An image dataset of circle markers is constructed automatically and training
manners of EllipseNet on the dataset are given. This construction method can auto-
matically generate a large amount of data for deep learning. Thus, EllipseNet can be
trained without manual annotations. At the same time, this dataset generating method
reduces the error caused by manual annotations. The proposed training loss function
takes account of the center and shape of the rotating bounding box, making the ellipse
region detection more accurate.
3. A circular marker is designed and camera localization is carried out based on
EllipseNet detector. Using the circular projective invariance, the analytical solution of
the camera pose is given. Experiments show that the proposed method is more robust to
far distance, illumination and viewpoint changes than the commonly used rectangular
markers.
修改评论