面向智能机器人的语义地图构建

Semantic map contains complete semantic descriptions (i.e. the semantic label of the instance objects) for the obstacles in the traditional obstacle map. Traditional SLAM algorithms seldom consider real complex dynamic scenarios. Currently, the semantic reconstruction methods can obtain the semantic information of scene while realizing the 3D reconstruction. However, these semantic information are mostly used for semantic feature matching to help estimate the position and orientation of the observation. At the same time, these semantic information is often pixel- or voxel-level, not in the form of instance object. However, these instantiated complete objects are the basic units in hierarchical (topological) map. For requirements of semantic mapping in complex dynamic scenes, this project focuses on the reliable separation for dynamic object, spatiotemporally consistent fusion of instance object and hierarchical semantic mapping, aiming to develop high efficiency and reliable method for dynamic scene reconstruction, explore the methods for space-time consistent fusion of instantiated object and hierarchical semantic mapping. We hope to make a key technology innovation in the semantic mapping for mobile robots under complex dynamic scenes, lay a foundation for the semantic intelligent interaction of mobile service robot and finally promote mobile service robot intelligent applications.

语义地图在传统的障碍物地图基础上还包含了障碍物的完整的语义信息（即实例对象级别的语义信息）。传统SLAM算法较少考虑真实复杂的动态场景。目前语义重建的研究工作，可以在实现场景三维重建的同时，获取场景的语义信息。但这些语义信息多用于语义特征匹配，辅助位置和姿态的估计；同时这些语义信息往往属于像素或体素级的语义，并未进一步形成关于物体的实例对象级语义表示，而这些实例化的完整对象正是组成层次化（拓扑）地图的基本单元。针对复杂动态场景下的地图构建需求，本项目围绕动态物体的可靠分离、实例对象的时空一致性融合、层次化语义地图的实时构建等核心科学问题，研究高效可靠的动态场景重建方法，探索时空一致的对象级语义完全化方法和层次化语义地图构建方法，实现机器人在复杂动态场景下语义地图构建的关键技术创新，为移动服务机器人的语义智能交互奠定基础，促进移动服务机器人的智能应用。

围绕复杂动态场景下智能机器人的场景重建、语义感知和语义地图构建，本项目重点研究了动/静态SLAM与场景重建和在线语义地图构建，建立了适应于复杂动态场景的高效的机器人语义地图构建系统，支持动态场景下高质量语义地图构建和增量式语义理解。有关技术通过与企业合作实现落地应用，取得了较好的效果。受本项目资助，累计在ACM Transactions on Graphics, IEEE Transactions on Visualization and Computer Graphics, Computational Visual Media, Computer Graphics Forum, Graphical Models, Computers & Graphics, CVPR, ECCV, ICRA等计算机图形学和视觉的重要期刊和会议上发表论文10篇,申请发明专利3项。