FreSTVO: Online Visual Odometry for Autonomous Vehicles With Frequency-Spatial-Temporal Constraints

Online visual odometry has been widely applied in autonomous vehicles for its fast adaptability to changing environments. Although existing algorithms successfully improve pose accuracy through data and computility-driven approaches, their rare focus on the thoughtful mining of temporal image sequence data may lead to the unsuitability for practical online utilization. Our method, FreSTVO, focuses on geometrical correspondences between motion information and high-dimensional features in autonomous driving to establish effective and comprehensive constraints among features. We design a Frequency-Domain Pose Consistency Refinement component to utilize phase analysis of bird's-eye-view projections to yield novel motion cues with rich gradient information for the constraint comprehensiveness. Additionally, Explicit Motion Feature Recoding is incorporated, which leverages geometric theories to extract high-dimensional motion features explicitly. Subsequently, only a single-layer ConvLSTM is used to extract motion context information efficiently. Further, frequency-spatial-temporal constraints between adjacent and interval poses are implemented in a self-supervised manner through Joint Spatial-Frequency Consistency Loss. Extensive experiments across the KITTI, Málaga, nuScenes, and CARLA datasets demonstrate that FreSTVO has average translation and rotation performance exceeding state-of-the-art methods up to 16.2% and 65.8%, respectively.