Multi-scale Direct Sparse Visual Odometry for Large-Scale Natural Environment

In this paper, we describe a multi-scale monocular direct sparse visual odometry (DSO) system to recover large-scale trajectories in unstructured natural environments in real time, while building a consistent metric map of the visited scenes. In contrast to the current state-of-the-art DSO system, the proposed method allows for more robust motion estimation and more accurate reconstruction in distant scenes by exploiting the characteristics of short- and long-range pixels, respectively. The long-range pixels, which are less sensitive to small camera translations, are used to initialize the camera rotation, so as to boost the tracking robustness in challenging natural environments. A multi-scale reconstruction framework is developed to recover short-range structure over successive frames, as well as the long-range structure over distant frames, hence allowing for a more consistent mapping precision. The reconstruction precision, the tracking accuracy, and the robustness of the proposed system are extensively evaluated with a publicly available vKITTI dataset, as well as the challenging Devon Island dataset, and Symphony Lake dataset. A detailed performance comparison between the proposed method and the state-of-the-art DSO system is presented.