Self-supervised Scale Recovery for Decoupled Visual-inertial Odometry

Accurate localization for intelligent robots remains a significant challenge, and self-supervised visual-inertial odometry (VIO) has emerged as a promising solution. However, existing self-supervised VIO works consider inertial information as the ordinary data input, losing its ability to recover absolute scales and ignoring the modality difference of acceleration and angular velocity in inertial data. In this letter, we present a novel self-supervised VIO framework that augments the odometry-related information implicit in inertial data. For the specific implementation, we propose a self-attention-based IMU network (IMUSAtt) to denoise the raw IMU data and then obtain the poses based on the denoised IMU data through an integrator. By constructing the pose consistency constraint between it and the visual-inertial fused pose, a Self-attention-based Scale Recovery (SSR) module is proposed to recover the absolute scale. Additionally, to avoid the interference of acceleration on rotation estimation, we design a Decoupled PoseNet (D-PoseNet) that employs different inputs and networks to learn rotation and translation. Odometry, scale, and depth evaluations on the KITTI odometry dataset and the Malaga dataset all reveal that our framework achieves state-of-the-art (SOTA) performance.