基于结构知识发现与推理的高适应性室内三维重建

Indoor space is the main place for people to live and work. Its three-dimensional reconstruction has always been an important research topic in the area of Information Processing, Remote Sensing and Mapping. The rapid development of sensing technology facilitates the collection of cost-effective 3D data, but the problem of complex scene and non-perfect data makes the transformation of raw data into a semantic, compact 3D model very challenging. An indoor space is an artificially constructed environment for a specific purpose, and its structural rules can be used to guide the reconstruction algorithm. The current indoor reconstruction research only adopts single-level, strong constraint rules, which are only applicable to specific scenarios, and the promotion ability is poor. This project will pioneer the mechanism of general structural knowledge in indoor scenes and explore a set of indoor 3D reconstruction schemes based on structural knowledge with extensive adaptability. This project uses multi-layer structure diagram and multi-layer Markov Random Field as tools to study the discovery, expression and reasoning mechanism of indoor scene structure knowledge, and automatically discover and flexibly apply indoor structure knowledge through machine learning. The applicant has accumulated a large number of research workers in the three-dimensional reconstruction of the house, laying a good foundation for the smooth implementation of this topic. The development of the project will lay the theoretical and technical foundation for the wide application of indoor three-dimensional models such as indoor navigation, BIM, and home improvement flexible customization.

室内空间是人们生活和工作的主要场所，其三维重建一直是信息处理及遥感测绘领域的重要研究课题。传感技术的快速发展为采集高性价比的三维数据提供了便利，但数据不够完善和室内结构复杂等困难为室内三维模型的语义化重建带来极大挑战。室内空间是为一定目的而人工建造的环境，其内在的结构规则可用于指导重建算法。当前室内重建研究只采用单层次的强约束规则，仅适用于特定场景。本项目将研究室内场景中通用的结构知识的机理，并探索一套基于结构知识的具备广泛适应能力的室内三维重建方案。本项目将以多层结构图和多层马尔科夫随机场为工具，研究室内场景的结构知识的发现、表达和推理机制，并探索自动发现并弹性应用室内结构知识的机器学习方法。申请人在房屋三维重建方面积累了较多的研究经验，为本课题的顺利实施奠定良好基础。课题的开展将为室内导航、建筑信息模型、家装柔性定制等室内三维模型的广泛应用奠定理论和技术基础。

室内空间是人们生活和工作的主要场所，其三维重建一直是信息处理及遥感测绘领域的重要研究课题。传感技术的快速发展为采集高性价比的三维数据提供了便利，但数据不够完善和室内结构复杂等困难为室内三维模型的语义化重建带来极大挑战。室内空间是为一定目的而人工建造的环境，其内在的结构规则可用于指导重建算法。当前室内重建研究只采用单层次的强约束规则，仅适用于特定场景。本项目研究室内场景中通用的结构知识的机理，并探索了一套基于结构知识的具备广泛适应能力的室内三维重建方案。本项目以多层结构图和多层马尔科夫随机场为工具，研究室内场景的结构知识的发现、表达和推理机制，并探索自动发现并弹性应用室内结构知识的机器学习方法。课题的开展将为室内导航、建筑信息模型、家装柔性定制等室内三维模型的广泛应用奠定理论和技术基础。项目完成预期成果，在国内外核心期刊和重要国际会议发表学术论文4 篇，其中在国内核心期刊发表论文1 篇，国际期刊和国际会议发表论文3篇。发表英文学术专著1本，申请国家发明专利2项，获得省部级科技进步奖1项，培养硕士6名。