Visual-Inertial Simultaneous Localization and Mapping with Point-Line Features

Visual-inertial simultaneous localization and mapping (VI-SLAM) integrates camera imagery and inertial measurements to estimate the trajectory of a sensor platform while constructing a three-dimensional representation of its environment. The inclusion of both point and line features enriches the geometric constraints available, improving performance in low-texture or dynamically lit scenes where point correspondences alone may be insufficient. Point features, such as corners and blobs, offer precise localisation but can suffer from sparsity; line features capture structural contours and edges, providing complementary information on orientation and scale. Modern VI-SLAM systems adopt tightly coupled frameworks in which pre-integrated inertial measurement unit (IMU) data and visual reprojection errors for points and lines are jointly minimised within a sliding-window optimisation. Advances in feature extraction, representation of 3D lines via Plücker coordinates or orthonormal parametrisations, and robust data association have driven improvements in accuracy and robustness. Practical real-world applications span autonomous vehicles navigating urban canyons, aerial drones performing indoor inspection in GPS-denied environments, and augmented-reality devices requiring centimetre-level pose stability. Current research focuses on enhancing global consistency through loop closure, integrating additional sensors such as LiDAR, and exploring learning-based approaches for more reliable feature detection and matching under adverse conditions.

Research from Nature Portfolio

No recent Nature Portfolio content available.

Research from all publishers

Recent developments outside the Nature Portfolio have concentrated on diverse system configurations and fusion strategies. A seminal monocular VI-SLAM approach employs a tightly coupled sliding-window optimisation that minimises both point and line reprojection errors alongside pre-integrated IMU residuals. By representing spatial lines with Plücker coordinates and an orthonormal basis, this system demonstrates enhanced trajectory estimation and map consistency compared with point-only schemes. In parallel, a LiDAR-visual-inertial odometry framework extends the classical VI-SLAM pipeline by projecting multi-frame LiDAR point clouds into the visual frame to establish depth correlations for extracted point-line features. A factor-graph formulation incorporating GNSS and loop-closure constraints yields globally consistent maps with reduced drift, achieving sub-decimetre accuracy indoors and a few metres outdoors. Earlier stereo visual-inertial navigation methods have also shown that the fusion of stereo imagery, point and line features, and inertial data through error-state extended and sigma-point Kalman filters can significantly improve robustness under rapid motion and non-linear dynamics, highlighting the enduring value of trifocal geometry and probabilistic filtering in VI-SLAM.

Visual-Inertial Simultaneous Localization and Mapping with Point-Line Features publication trend

The graph below shows the total number of articles in visual-inertial simultaneous localization and mapping with point-line features across all publications each year (not limited to Nature Index journals).

Technical terms

Visual-Inertial SLAM: simultaneous localisation and mapping using camera data and inertial measurements to estimate pose and reconstruct the environment.

Point Features: distinct, localised image elements (corners or blobs) used for establishing correspondences between frames.

Line Features: linear image structures representing edges or boundaries, employed to augment geometric constraints in mapping.

IMU Pre-integration: technique to summarise high-rate inertial measurements between keyframes, reducing computational complexity.

Plücker Coordinates: algebraic six-parameter representation of 3D lines facilitating optimisation in SLAM back-ends.

Sliding-Window Optimisation: framework that refines a limited recent sequence of poses and landmarks to balance accuracy and efficiency.

Factor Graph: probabilistic graphical model linking sensor observations and state variables for joint inference in sensor fusion.

References

1. PL-VIO: Tightly-Coupled Monocular Visual–Inertial Odometry Using Point and Line Features. Sensors (2018).
2. Tightly-Coupled Stereo Visual-Inertial Navigation Using Point and Line Features. Sensors (2015).
3. LiDAR-Visual-Inertial Odometry Based on Optimized Visual Point-Line Features. Remote Sensing (2022).