Real-time 3D Semantic Instance Segmentation SLAM

Abstract

The interaction between mobile robot and environment is the requirement of intelligent robot. On the basis of real-time localization, the robot needs to reconstruct higher level entities during mapping. This paper proposes a real-time 3D semantic segmentation SLAM algorithm, which gradually segments the instance objects in the environment through the information in the field of view. The algorithm in this paper segments instances by the nature of the normal vector of the object surface, and uses the object recognition CNN to obtain semantic labels for semantic segmentation of key frames. In the construction of information association and 3D instance, this paper proposes a low-cost information association octree model, which realizes semantic information unification and instance merging in grid space. The local common view relationship and recursive update method are used to avoid the average semantic segmentation time increasing over time. The proposed algorithm is tested on public data sets, and the experimental results show that the proposed algorithm has reliable accuracy and good real-time performance.