POPL-SLAM: A Pose-Only Representation-Based Visual-Inertial SLAM With Point and Structural Line Features

The geometric regularity of man-made structures imposes distinct constraints on Visual-Inertial Simultaneous Localization and Mapping (VI-SLAM) systems. To effectively leverage this geometric information, we integrate distance and angular data from linear structural features to enforce additional constraints on point-based VI-SLAM. To mitigate the computational overhead of high-level features, we propose POPL-SLAM, which adopts a pose only approach to implicitly model points and structural lines. First, we develop a progressive vanishing point estimation technique lever aging IMU preintegration priors. By formulating a non-orthogonal constraint optimization, this technique efficiently estimates vanishing points without per-frame recalculations. Second, we introduce a pose-only framework for a point-line SLAM system, implicitly modeling points and lines to formulate constraints. This approach eliminates linearization errors during optimization and markedly improves computational efficiency. For structural line features, we concurrently incorporate Euclidean and angular distance errors to comprehensively capture positional and orientational information. Extensive experiments on public real-world datasets demonstrate that POPL-SLAM outperforms state-of-the-art point-line VI-SLAM systems in accuracy, efficiency, and robustness.