Robust underwater SLAM with dual-channel disparity prediction and yaw-constrained multisensor fusion

Underwater Simultaneous Localization and Mapping (SLAM) remains a challenging problem due to severe light attenuation, scattering-induced image degradation, and the unavailability of Global Positioning System (GPS) signals. As a result, both visual-only and inertial-only odometry typically suffer from drift accumulation and unreliable data association in low-texture underwater environments. To overcome these challenges, we develop a robust multi-sensor underwater SLAM framework that integrates monocular vision, an Inertial Measurement Units (IMUs), a pressure sensor, and a magnetometer. At the front end, a dual-channel disparity prediction strategy enhances feature correspondence by jointly leveraging horizontal motion inferred from IMU pre-integration and vertical displacement estimated from pressure readings. A yaw-constrained estimation module further incorporates magnetometer-based heading cues to improve yaw observability under weak rotational excitation. All sensor modalities are fused in a factor graph optimization back-end to ensure globally consistent state estimation. Experiments on the AQUALOC dataset show that the proposed system reduces trajectory drift by over 80%, decreases depth error by more than 75%, and lowers yaw error by approximately 60% compared to a visual–inertial SLAM baseline, demonstrating its robustness and applicability in real underwater environments.