Awareness-Guided Self-Supervised Visual Odometry via Fusion of Pseudo-Depth and Residual-Flow Cues
Abstract
Highlights What are the main findings? Pseudo-depth and residual flow provide complementary geometric and motion cues for identifying unreliable cross-frame observations caused by non-observable regions and independently moving objects. Incorporating awareness of reliable cross-frame observations into self-supervised pose learning at both the input and loss levels improves ego-motion estimation and reduces trajectory errors. What are the implications of the main findings? Awareness of the reliability of cross-frame observations provides a more effective basis for self-supervised pose learning than treating all image regions as equally reliable. Combining complementary geometric and motion information offers a practical way to improve the accuracy of self-supervised visual odometry. Abstract Self-supervised monocular visual odometry (VO) commonly relies on photometric consistency and rigid-scene assumptions to learn camera motion, making pose estimation vulnerable to unreliable cross-frame observations caused by limited geometric observability and independently moving objects. To address this problem, we propose an awareness-guided self-supervised visual odometry framework that emphasizes reliable cross-frame observations for pose learning. Pseudo-depth is exploited to characterize geometric observability and identify regions that are unreliable for cross-frame reconstruction, while residual flow is employed to capture motion inconsistency introduced by independently moving objects. The resulting observable static regions are incorporated into pose estimation at both the input and loss levels, enabling unreliable observations to be suppressed before feature extraction and during optimization. Qualitative experiments on KITTI and Cityscapes verify the effectiveness of pseudo-depth and residual flow in extracting geometric observability and dynamic-motion information across different driving scenes. Quantitative experiments on the KITTI Odometry dataset show that the proposed method achieves average relative translation and rotation errors of 6.04% and 2.42°/100 m, together with an ATE of 0.88 m and an ARE of 0.20°, demonstrating improved ego-motion and trajectory estimation.