Occlusion-Aware Monocular Visual Odometry for Robust Trajectory Tracking

Abstract

Hybrid monocular visual odometry, which combines the advantages of learning-based and geometry-based methods, has gained widespread attention for its superior robustness and accuracy. Recent works use networks to predict optical flow to establish pixel correspondences between images while retaining a geometry-based nonlinear optimization backend. However, existing methods primarily rely on local features to refine optical flow. A critical limitation of these methods is that the loss of local features due to occlusion can introduce significant uncertainty, degrading the performance of visual odometry. To enhance the system's robustness, we propose an occlusion-aware monocular visual odometry that aggregates both spatial and temporal features, effectively leveraging global information to reduce the impact of occlusion. Our method consists mainly of a Spatial Feature Aggregation (SFA) module and a Temporal Feature Aggregation (TFA) module. SFA models image self-similarity, utilizing visible regions to guide optical flow estimation in occluded regions. Meanwhile, TFA captures dynamic variations across consecutive frames, enhancing the model's understanding of motion trends. Experiments demonstrate that our method achieves SOTA performance on multiple benchmarks. Notably, compared to the baseline, it reduces the average Absolute Trajectory Error (ATE) on the TartanAir test split by 25.71%.