An RGB-D Semantic Map Building and Global Localization Method

Aiming at the shortcomings of indoor visual SLAM system such as lack of environment awareness, sparse map construction, low global positioning accuracy and poor robustness, this paper proposes a semantic map building and global positioning method based on visual semantic descriptors and dense point cloud map. Abstract semantic descriptors are extracted and keyframe dense point cloud map is built by real-time object detection of color images. Global positioning adopts the off-line positioning method combining coarse and fine tuning. By quickly comparing the similarity between descriptors in the map set, the most similar reference keyframe is selected from the keyframe map, and the pose is obtained as the rough estimation result. The iterative calculation is continued in the dense point cloud map to complete the off-line positioning. In this paper, an available semantic mapping and positioning system is constructed and tested on the public RGB-D sequence dataset. The result shows that the proposed method can generate high-quality indoor point cloud maps and also finish global localization with higher accuracy and better robustness than the classical algorithm.