Vanishing Points Detection with Line Segments of Gaussian Sphere

Abstract

Depth estimation using vanishing points is important in computer vision and has been widely used in various applications such as robotics and autonomous driving. A vanishing point is a point in the image where parallel lines appear to converge to a single point in 3D space. The detection of vanishing points in images plays a crucial role in estimating the depth of a scene. However, the accuracy of vanishing point detection is often affected by noisy or unconverged line segments detected by the line detectors. The problem with using line detectors is that they can produce noisy or unconverged line segments, leading to a decrease in the accuracy of vanishing point detection. Therefore, it is important to develop a method to extract accurate vanishing points from noisy line segments. This paper proposes an algorithm to detect vanishing points by projecting line segments to Gaussian sphere. The proposed method follows the steps: (1) Line segment detection by Mobile LSD [1], (2) Classifying line segments based on their angle, and (3) Projecting the classified line segments onto a Gaussian sphere and converting them into a 2D image. In the context of this transformed 2D image, it is postulated that the vanishing point corresponds to the region in which the line segments exhibit the greatest degree of overlap. In other words, the proposal does not need to compute all intersection points from line segments and recognize the vanishing points from the intersection points. The proposal can not only improves the accuracy but also reduces the time in vanishing points detection by the experiments.