Semantics-guided reconstruction of indoor navigation elements from 3D colorized points

摘要

The increasing availability of both indoor positioning services and sensors for 3D data capture, such as RGB-D sensors, allows the provision of indoor spatial information services for indoor localization-based applications. To efficiently realize these services, the indoor information and the relationships between indoor spaces are required. The recently released Indoor Geography Markup Language (IndoorGML) attempts to represent and exchange geo-information for modeling topology and semantics of indoor spaces. However, it is still challenging to map indoor space features to the IndoorGML-encoded navigation network model directly from colorized 3D points. Therefore, we propose a semantics-guided method for indoor navigation element reconstruction from RGB-D sensor data. First, a hierarchical indoor scene interpretation framework is used for robustly recognizing the architecture structures and doors, respectively. In the developed hierarchical structure, a graph convolutional network-based architectural structure recognition method is adopted to deduce the long-range interactions among primitives for describing the rich physical relationships in the real world. Its output is the produced initial annotated results, from which doors as the common openings are further detected using a U-Net-based door recognition method. This enables to effectively provide the semantic guidance for the cellular representation of the indoor space and its topological reconstruction. Second, an adaptive architectural structure-guided room segmentation method is developed by combining distance transform and watershed segmentation to determine cellular spaces according to the definition in IndoorGML. Third, taking the different states of doors into consideration, a door-guided topological relationship reconstruction method is proposed to achieve the network graph representation of indoor environments. In this context, a simulated door model is designed to correct and update the true position of a door leaf, and a virtual door is defined to optimize the topological analysis. As a consequence, an IndoorGML-encoded navigation network model is generated, which can be used as the base for indoor navigation applications independent of the platform. Experiments are performed on the public Stanford large-scale 3D Indoor Spaces Dataset to verify the robustness and effectiveness of the proposed method both qualitatively and quantitatively. Results indicate the capability of the proposed method in automatically reconstructing indoor navigation elements of Manhattan-world indoor environments from RGB-D sensor data.