| Nowadays indoor scene labeling has become a hot topic in the reform of machine learning, artificial intelligence, computer vision, etc. Its use enables mobile robots to perform scene understanding and their tasks more reliably, flexibly, and efficiently. Scene labeling in3D point cloud faces many problems such as the segmentation of point cloud, the extraction and description of point cloud’s features, the modeling for learning algorithm and inference, etc. A significant amount of work has been done in3D semantic labeling already. However, the difficulties of the instability of segmentation, the insufficiency of properties’description, the locality of contextual constraints still exist.Based on the previous related work, this paper puts forward a systematic approach for indoor scene3D semantic labeling. The input is multiple Kinect RGBD point clouds of an indoor scene. A mobile robot with a Kinect can perform3D mapping while labeling. As the data captured by a Kinect is usually noisy and incomplete, fusion model based region growing algorithm is presented for segmentation of each point cloud frame. Then each segment will be represented using fusion features including visual appearance and shape cues. The co-occurrence relationships and geometric relationships among segments from consecutive point cloud frames are modeled by the Markov Random Filed (MRF). Unlike the traditional local contextual constrains based on one frame, a global ones is computed iteratively using Iterative MRF along with3D mapping process simultaneously.A comparison experiment was conducted between our labeling approach and a traditional labeling method on a public3D dataset. This dataset contains523D indoor scenes with about550views. The results indicate the effectiveness of our approach. Finally our approach was experimented in a real indoor environment along with3D mobile mapping process. The results show that our approach can accomplish the labeling task robustly and effectively. |
PDF Full Text Request