A hybrid formulation for monocular visual SLAM

Simultaneous Localization and Mapping (SLAM) is the problem of localizing a sensor in a map that is built online. SLAM technology can enable robot localization in unknown environments by processing onboard sensors and therefore not relying on external infrastructure. A map allows an agent to continually localize itself in the same environment without accumulating drift, in contrast to odometry approaches where incremental motion is integrated over time. Such a technology is critical for the navigation of service robots and autonomous vehicles, or to localize a user in virtual or augmented reality applications. Typically visual SLAM system can be broadly categorized as being Direct or Indirect. Indirect methods build on an alternative image representation based on visual feature descriptors and compute geometric residuals via triangulation while, on the other hand, direct methods skip the pre-processing step and deal with image pixels directly to generate photometric residuals. In recent years big improvements have been achieved in both directions, but few works exist regarding hybrid approaches. This thesis explores a hybrid formulation for visual SLAM, in which direct andindirect approaches are fused together exploiting the complementary properties of both paradigms. We propose a monocular algorithm where the bundle adjustment problem is solved with a coupling of photometric (Direct) and geometric (Indirect) measurements inside a joint multi-objective optimization framework. Quantitative results have been computed using the popular monocular visual SLAM benchmark from TUM University [1] and they show that the proposed formulation is capable of handling large inter-frame motions, outputs an Indirect both paradigms. We propose a monocular algorithm where the bundle adjustment problem is solved with a coupling of photometric (Direct) and geometric (Indirect) measurements inside a joint multi-objective optimization framework. Quantitative results have been computed using the popular monocular visual SLAM benchmark from TUM University [1] and they show that the proposed formulation is capable of handling large inter-frame motions, outputs an Indirect 3D map which can be further used for loop-closing or relocalization, achieves sub-pixel precision accuracy like direct methods while still running in real-time.