Mamba model guided deep visual-inertial odometry

Visual-inertial odometry (VIO) aims to predict trajectories by fitting data from cameras and inertial measurement unit (IMU). In recent years, deep learning-based VIO has made significant progress. However, several challenges persist, such as the loss of key visual features during sudden linear acceleration and the efficient fusion of visual and inertial information. To address these challenges, we propose a Mamba model guided deep visual-inertial odometry algorithm, named MamVIO. Specifically, we carefully designed a module for extracting visual features, which dynamically adjusts spatiotemporal receptive field to capture the dynamic variations of key features along the temporal dimension and aggregates information from adjacent frames. Furthermore, to effectively combine visual and inertia information, we design a Mamba model based visual-inertial fusion module, which extends the Mamba model to support dual input. Extensive evaluations on the public odometry benchmark demonstrate that our algorithm achieves competitive performance, increasing average translational accuracy by 9.56% and average rotational accuracy by 8.55%, respectively, compared to baseline method.