GD-MVO: Monocular Geometric Visual Odometry Using Image Depth Prediction

Traditional geometric methods estimate camera motion trajectories by analyzing image feature points or pixel information, demonstrating robust performance in certain scenarios. However, these approaches struggle in low-texture and highly dynamic environments. In contrast, end-to-end neural networks can directly predict camera motion trajectories from consecutive image frames, eliminating the need for complex feature extraction and matching processes. Nonetheless, the accuracy of this end-to-end approach still has a significant gap compared to traditional methods. We designed GD-MVO, a monocular visual odometry system that combines the strengths of two approaches. Our system uses both monocular and stereo image pairs to train a deep neural network for depth prediction, then integrates this depth information with geometric features extracted from sequential images to achieve more accurate visual positioning. In this work, we propose a) a novel integration of geometric methods and deep learning for robust visual odometry; b) an effective strategy to mitigate the impact of dynamic feature points on prediction outcomes; c) a fly-out mask technique to resolve the fly-out padding issue in monocular depth prediction networks. Experimental results show that our GD-MVO not only exhibits excellent stability in high-dynamic scenes but also maintains reliable running performance on our specially customized low-texture dataset. In the KITTI odometry benchmark, GD-MVO outperforms both DF-VO and TSformer-VO across all metrics. Most impressively, it achieves a translation error of merely 1.778%, surpassing DF-VO (6.747%) and RAUM-VO (3.676%).