SP-VO: RGB-D Visual Odometry Using Static Parts Toward Dynamic Environments

Estimating visual odometry in dynamic environments is a challenging problem, as features of moving objects prevent accurate image matching. Typical approach to deal with dynamic environments is to remove features of moving objects. However, this results in a lot of information loss and makes it impossible to use static parts of moving objects that could be fully utilized for image matching process. To address this issue, we propose a geometrical inference approach that utilizes the static parts of moving objects and background to achieve accurate feature matching. Moreover, we propose the concept of matching confidence that is calculated by comparing the squared residual motion likelihood with the chi-square distribution. For each frame, the geometric model and the semantic model are selected according to the proposed confidence, so that visual odometry could be estimated more accurately. Our algorithm was evaluated on RGB-D datasets, including dynamic environments. The results show better performance than prior algorithms.