Point-line visual-inertial odometry with length-filtered and collinearity-labeled line segments

This is the authors' abstract. We don't add key points for this paper.

Reliable extraction of line segments from complex scenes has been enabled by the line segment detector (LSD) algorithm, which provides accurate geometric information for various visual perception applications. However, LSD often extracts line segments that are either excessively short or overly long, increasing computational complexity and degrading feature matching accuracy. Although many methods focus on eliminating redundant short lines, the challenges of overly long line segments remain largely unexplored. To address these issues, this paper proposes an enhanced LSD-based method that integrates short-segment elimination and long-segment truncation strategies. The resulting line segments are constrained within a reasonable length range, effectively balancing geometric information with feature matching robustness. In addition, a collinearity-based matching method is introduced to maintain the geometric continuity of truncated segments and improve inter-frame feature consistency. Furthermore, the proposed method is integrated into a point-line visual-inertial odometry system by fusing data from an inertial measurement unit, significantly improving system robustness and accuracy, especially in low-texture environments. Extensive experiments on public datasets and real-world scenarios show that the proposed method achieves competitive overall performance, particularly in structured and low-texture environments.