Beyond Learning: Back to Geometric Essence of Visual Odometry via Fusion-Based Paradigm

Camera-based motion estimation, also called visual odometry (VO), is a technique applied in the area of autonomous driving and simultaneous localization and mapping (SLAM), which is gaining huge popularity due to its simplicity and use of limited resources in localization. In this article, we propose to consider the depth and optical flow estimation as a joint learning task and then obtain ego motion based on geometric constraints. To be specific, we construct a network to learn dense optical flow and depth map simultaneously, where the scale of the prediction is aligned by an additional velocity supervision signal. Then, we formulate the uncertainty of the two main geometric methods and leverage covariance intersection (CI) fusion, allowing the ego motion to be efficiently predicted in an automodeling and compensation way, without manual intervention, postprocessing, or pose network training utilized before. Compared with recent approaches to similar tasks, our method requires neither complicated feature descriptor calculations nor expensive supplemental sensors. Moreover, our model flexibly adapts to both unsupervised and weakly supervised learning, which improves the feasibility of the framework in diverse settings. We conduct extensive experiments on several datasets representing driving and walking scenarios, and the results demonstrate a promising performance of the proposed method to state-of-the-art monocular ones.