Systematic Analysis of Direct Sparse Odometry

Abstract

In the field of robotics and autonomous driving, the camera as a sensor gets more and more important, as the camera is cheap and robust against environmental influences. One challenging task is the localization of the robot on an unknown map. This leads to the so-called Simultaneous Localization and Mapping (SLAM) problem. For the Visual SLAM problem, a plethora of algorithms was proposed in the last years, but the algorithms were rarely evaluated regarding the robustness of the approaches. This contribution motivates the systematic analysis of Visual SLAMs in simulation by using heterogeneous environments in Blender. For this purpose, three different environments are used for evaluation ranging from very low detailed to high detailed worlds. In this contribution, the Direct Sparse Odometry (DSO) is evaluated as an exemplary Visual SLAM. It is shown that the DSO is very sensitive to rotations of the camera. In addition, it is presented that if the scene does not provide sufficient clues about the depth, an estimation of the trajectory is not possible. The results are complemented by real-world experiments.