Effective and Efficient Line Segment Detection for Visual Measurement Guided by Level Lines
Xinyu Lin,Yingjie Zhou,Yipeng Liu,Ce Zhu
DOI: https://doi.org/10.1109/tim.2023.3328094
IF: 5.6
2023-01-01
IEEE Transactions on Instrumentation and Measurement
Abstract:Line segment detection is the basis for various visual measurement tasks. Numerous methods have been proposed to detect line segments from images, and edge-fitting-based ones have gained significant attention because of their remarkable detection efficiency. However, most edge-fitting-based methods primarily rely on gradient magnitude for edge detection and edge coordinates for line segment fitting, neglecting the importance of considering gradient orientation, which may reduce their effectiveness. In addition, most of them require the least squares for line segment fitting, involving the computationally inefficient squaring operation. To solve the above issues, this study proposes an effective and efficient line segment detection (E2LSD) algorithm based on two new findings regarding the level line of edge points, i.e., the unit vector orthogonal to the corresponding gradient orientation. 1) Utilizing double consistent constraints of both coordinates and level lines of edge points to fit line segments results in a more effective line segment detection than those relying on a single consistent constraint of coordinates. 2) Decoupling line segment orientation and position, followed by fitting them separately using level lines and coordinates of edge points, results in a computationally efficient line segment detection approach. It is more computationally efficient than those directly fitting line segments in the least squares sense based on the coordinates of edge points. In the E2LSD algorithm, edges are drawn with the guideline of level lines to improve accuracy. Numerical experiments based on natural and synthetic datasets showed that the E2LSD algorithm outperforms existing state-of-the-art (SOTA) methods regarding both effectiveness and computational efficiency. The E2LSD algorithm has also successfully been employed in a visual measurement system regarding feature-based visual localization. The code of the E2LSD algorithm will be publicly available at https://github.com/roylin1229/E2LSD .