This paper presents a novel semantic mapping approach, Recurrent-OctoMap, learned from long-term 3D Lidar data. Most existing semantic mapping approaches focus on improving semantic understanding of single frames, rather than 3D refinement of semantic maps (i.e. fusing semantic observations). The most widely-used approach for 3D semantic map refinement is a Bayes update, which fuses the consecutive predictive probabilities following a Markov-Chain model. Instead, we propose a learning approach to fuse the semantic features, rather than simply fusing predictions from a classifier. In our approach, we represent and maintain our 3D map as an OctoMap, and model each cell as a recurrent neural network (RNN), to obtain a Recurrent-OctoMap. In this case, the semantic mapping process can be formulated as a sequence-to-sequence encoding-decoding problem. Moreover, in order to extend the duration of observations in our Recurrent-OctoMap, we developed a robust 3D localization and mapping system for successively mapping a dynamic environment using more than two weeks of data, and the system can be trained and deployed with arbitrary memory length, benefiting from a "napping" mechanism. We validate our approach on the ETH long-term 3D Lidar dataset [1]. The experimental results show that our proposed approach outperforms the conventional "Bayes update" approach, and the proposed long-term memory model outperforms a standard short-term memory model.

Robot-assisted endobronchial intervention requires accurate localisation based on both intra- and pre-operative data. Most existing methods achieve this by registering 2D videos with 3D CT models according to a defined similarity metric with local features. Instead, we formulate the bronchoscopic localisation as a learning-based global localisation using deep neural networks. The proposed network consists of two generative architectures and one auxiliary learning component. The cycle generative architecture bridges the domain variance between the real bronchoscopic videos and virtual views derived from pre-operative CT data so that the proposed approach can be trained through a large number of generated virtual images but deployed through real images. The auxiliary learning architecture leverages complementary relative pose regression to constrain the search space, ensuring consistent global pose predictions. Most importantly, the uncertainty of each global pose is obtained through variational inference by sampling within the learned underlying probability distribution. Detailed validation results demonstrate the localisation accuracy with reasonable uncertainty achieved and its potential clinical value.

In this paper, a novel Pixel-Voxel network is proposed for dense 3D semantic mapping, which can perform dense 3D mapping while simultaneously recognizing and labelling the semantic category each point in the 3D map. In our approach, we fully leverage the advantages of different modalities. That is, the PixelNet can learn the high-level contextual information from 2D RGB images, and the VoxelNet can learn 3D geometrical shapes from the 3D point cloud. Unlike the existing architecture that fuses score maps from different modalities with equal weights, we propose a softmax weighted fusion stack that adaptively learns the varying contributions of PixelNet and VoxelNet and fuses the score maps according to their respective confidence levels. Our approach achieved competitive results on both the SUN RGB-D and NYU V2 benchmarks, while the runtime of the proposed system is boosted to around 13 Hz, enabling near-real-time performance using an i7 eight-cores PC with a single Titan X GPU.

This paper presents an interactive approach to the construction of a grid-semantic map for the navigation of service robots in an indoor environment. It is based on the Robot Operating System (ROS) framework and contains four modules, namely Interactive Module, Control Module, Navigation Module and Mapping Module. Three challenging issues have been focused during its development: (i) how human voice and robot visual information could be effectively deployed in the mapping and navigation process; (ii) how semantic names could combine with coordinate data in an online Grid-Semantic map; and (iii) how a localization–evaluate–relocalization method could be used in global localization based on modified maximum particle weight of the particle swarm. A number of experiments are carried out in both simulated and real environments such as corridors and offices to verify its feasibility and performance

Recognizing material categories is one of the core challenges in robotic nuclear waste decommissioning. All nuclear waste should be sorted and segregated according to its materials, and then different disposal post-process can be applied. In this paper, we propose a novel transfer learning approach to learn boundary-aware material segmentation from a meta-dataset and weakly annotated data. The proposed method is data-efficient, leveraging a publically available dataset for general computer vision tasks and coarsely labeled material recognition data, with only a limited number of fine pixel-wise annotations required. Importantly, our approach is integrated with a Simultaneous Localization and Mapping (SLAM) system to fuse the per-frame understanding delicately into a 3D global semantic map to facilitate robot manipulation in self-occluded object heaps or robot navigation in disaster zones. We evaluate the proposed method on the Materials in Context dataset over 23 categories and that our integrated system delivers quasi-real-time 3D semantic mapping with high-resolution images. The trained model is also verified in an industrial environment as part of the EU RoMaNs project, and promising qualitative results are presented. A video demo and the newly generated data can be found at the project website¹ (Supplementary Material).