A unified adaptive weighting framework for GNSS/visual/inertial odometry with deep-learned point-line features
Abstract
Visual-inertial odometry often suffers from performance degradation in challenging environments due to the failure of handcrafted features and fixed-weight fusion schemes. This paper proposes a tightly-coupled global navigation satellite system (GNSS)/visual/inertial odometry system with two core advances: a deep learning-based point-line feature front-end and a geometry-aware adaptive weighting strategy that dynamically adjusts the contribution of visual and GNSS observations. The front-end employs SuperPoint with SuperGlue for robust point extraction and matching, and the Richer Convolutional Features network for enhanced edge detection to obtain reliable line segments. For the back-end, we derive analytically the sensitivity of point reprojection error to depth and of line reprojection error to the line’s orientation relative to the camera axis, and construct adaptive weight functions that replace the conventional fixed covariances. Crucially, this adaptive weighting strategy is extended to GNSS measurements by incorporating the dilution of precision, carrier-to-noise ratio, and multipath anomaly detection, thus enabling a unified framework where every observation is weighted according to its instantaneous geometric reliability. Extensive experiments on the EuRoC MAV dataset, and self-collected campus sequences demonstrate that the proposed method achieves superior accuracy and robustness in most standard and moderately degraded scenarios compared to representative baseline methods. On the EuRoC MAV dataset, the proposed unified framework reduces the average Root-Mean-Square of the absolute trajectory error by 8.4% relative to fixed-weight PL-VINS, while the deep feature front-end further improves robustness and trajectory accuracy in low-light and low-texture environments. In GNSS-degraded urban canyons, the adaptive GNSS weighting mechanism suppresses the influence of unreliable satellite observations without manual tuning, leading to more stable and reliable localization.