Accurate stereo visual-inertial SLAM with rapid initialization and map point sparsification

Simultaneous Localization and Mapping (SLAM) technology is crucial in applications like Augmented Reality (AR) and Virtual Reality (VR), providing precise spatial positioning and scene perception. This paper proposes an enhanced Visual-Inertial Odometry (VIO) initialization strategy based on ORB-SLAM3, which independently estimates gyroscope bias and decouples rotation and translation, significantly reducing VIO initialization time and improving trajectory accuracy. We introduce an improved inertial measurement units (IMU) preintegration formula based on autohomomorphism of SE₂(3) to enhance tracking performance under rapid rotation and extended integration periods. To address redundant map information, a novel map point elimination strategy is employed, prioritizing pose visibility and spatial diversity. Experimental results on the EuRoC dataset show that our method shortens the VIO initialization start time shortening, reduces the number of map points, and reduces the tracking time and local map building time, confirming the effectiveness of our approach compared to state-of-the-art methods.