Object-Oriented 3D Semantic Mapping Based on Instance Segmentation

Abstract

Service robots gain both geometric and semantic information about the environment with the help of semantic mapping, providing more intelligent services. However, a majority of studies for semantic mapping thus far require priori knowledge 3D object models or maps with a few object categories that neglect separate individual objects. In view of these problems, an object-oriented 3D semantic mapping method is proposed by combining state-of-the-art deep-learning-based instance segmentation and a visual simultaneous localization and mapping (SLAM) algorithm, which helps robots not only gain navigation-oriented geometric information about the surrounding environment, but also obtain individually-oriented attribute and location information about the objects. Meanwhile, an object recognition and target association algorithm applied to continuous image frames is proposed by combining visual SLAM, which uses visual consistency between image frames to promote the result of object matching and recognition over continuous image frames, and improve the object recognition accuracy. Finally, a 3D semantic mapping system is implemented based on Mask R-CNN and ORB-SLAM2 frameworks. A simulation experiment is carried out on the ICL-NUIM dataset and the experimental results show that the system can generally recognize all the types of objects in the scene and generate fine point cloud models of these objects, which verifies the effectiveness of our algorithm.