PML-SLAM: Optimizing and Enhancing Visual SLAM with Point-to-Line Matching

Visual feature matching and state estimation are core issues in Simultaneous Localization and Mapping (SLAM) systems. Traditional methods typically extract and match feature points and line segments separately in images to build robust visual odometry. However, this decoupled approach overlooks the intrinsic connection between points and line segments, resulting in redundant computation and suboptimal performance. This paper presents a novel approach that derives line segment matches directly from point matches using our proposed geometric constraint algorithm. By leveraging the implicit relationships between matched points, we eliminate the need for explicit extraction and matching of the line descriptions, addressing a fundamental inefficiency in SLAM systems. Experimental results demonstrate that this method significantly reduces the computational resource consumption while maintaining or even improving stability and accuracy. In the EuRoC and TUM-VI dataset, our approach decreases processing time by 67% compared to the SLAM system, which extracts and matches feature points and line segments separately, while improving localization accuracy by 18% compared to the baseline method. The integration of our unified feature matching approach into existing SLAM frameworks demonstrates its broad applicability. This research provides a new perspective for the design of visual SLAM systems by challenging conventional separate processing of different feature types and offers valuable insights for other feature-matching-based computer vision tasks.