【作者】 王志宇；

【导师】 丁伟利； 黄向生；

【作者基本信息】 燕山大学 ， 人工智能（专业学位）， 2025， 博士

【摘要】 随着服务机器人在家庭、餐饮、医疗等场景中的不断应用,如何在复杂室内环境中实现高精度、强鲁棒的定位与建图成为亟待解决的关键问题。当前主流基于特征的视觉同步定位与建图(Simultaneous Localization and Mapping,SLAM)方法多以点、线或面等单一几何元素构建约束,其结果易受噪声、遮挡与误检等因素影响,导致系统在复杂室内环境下的定位可靠性较低。为解决上述问题,本文以提升服务机器人在真实室内环境中的定位精度与稳定性为目标,基于室内人造环境中普遍存在的平行线簇、平行平面和正交平面等规则几何结构所具有的稳定且一致的关联约束信息,展开多粒度特征融合SLAM框架设计以及动态环境下服务机器人的鲁棒定位方法研究,具体内容如下:首先,针对现有几何结构检测方法在复杂室内场景中鲁棒性不足的问题,提出了基于多尺度几何约束的多粒度几何结构检测方法。该方法综合利用角点、平行线结构和平面结构的几何结构,进行点-平行线-平面结构检测。并基于投影重叠、一致性与近似垂直生长准则,建立多场景评测数据集与量化指标体系,实现了对平行结构的高效检测和科学评估,解决了现有平行线结构检测依赖简单约束、难以适应复杂环境的问题,为后续SLAM特征构建奠定了可靠的结构基础。其次,针对静态室内场景中重复纹理与摆放物体导致线条杂乱、易引发常规点-线SLAM漂移与失效的问题,基于所提平行结构检测结果,提出了一种点-平行线融合SLAM方法。该方法通过同步构建点与平行线特征,并在全局与局部层面充分利用平行线约束,以提升定位精度与系统稳定性。然而,在极弱纹理或复杂光照条件下,系统仍面临特征约束不足的问题。为此,进一步提出点-平行线-平面融合的多粒度特征SLAM方法,在保持中层结构约束优势的同时,通过平面参数化建模与局部曼哈顿约束增强环境几何表达能力。实验结果表明,平行线特征的引入显著提升了系统在多类室内环境下的定位精度与鲁棒性,尤其在视觉复杂但具有部分规则几何布局的场景中表现更优;多粒度特征融合进一步提升了系统在复杂静态服务场景中的精度与建图完整性。接下来,针对真实服务场景中移动目标对系统定位与建图的干扰问题,提出了融合动态建模机制的多粒度特征SLAM方法。通过构建特征动态概率模型,结合观测与时序传播概率,为多粒度特征分配动态属性,从而实现动态特征的检测与剔除。该机制仅利用多粒度静态特征进行定位与建图,在保持多几何约束优势的同时可整体提升系统的动态鲁棒性。实验结果表明,动态特征提取策略可保障服务机器人在人员密集、开放式食堂等复杂动态场景中保持稳定跟踪。最后,本文面向家庭、办公室与低纹理走廊等典型服务场景,自建了服务机器人实验平台,并在多个真实场景中开展了系统应用验证,证明了提出的多粒度几何特征SLAM方法在定位精度、鲁棒性及环境适应性方面均优于现有方法,并展现出良好的普适性与实际应用潜力。更多还原

【Abstract】 With the rapid deployment of service robots in domestic,catering,healthcare,and other indoor scenarios,achieving accurate and robust localization and mapping in complex indoor environments has become a critical challenge.Most existing feature-based visual simultaneous localization and mapping(SLAM)methods construct geometric constraints using a single type of primitive,such as points,lines,or planes.These approaches are often vulnerable to noise,occlusions,and false detections,which significantly degrades localization reliability in cluttered and structurally complex indoor environments.To address these limitations,this thesis aims to improve the localization accuracy and stability of service robots operating in real indoor environments.Leveraging the stable and consistent geometric relationships inherent in man-made indoor structures,including clusters of parallel lines,parallel planes,and orthogonal planar layouts,we investigate the design of a multi-granularity feature fusion SLAM framework together with robust localization strategies for dynamic service environments.The main contributions are summarized as follows.First,to overcome the limited robustness of existing geometric structure detection methods in complex indoor scenes,a multi-granularity geometric structure detection approach based on multi-scale geometric constraints is proposed.The method jointly exploits corner points,parallel line structures,and planar structures to perform point-parallel line-plane detection.By incorporating projection overlap,geometric consistency,and approximate vertical growth criteria,a multi-scenario benchmark dataset and quantitative evaluation protocol are established,enabling efficient detection and rigorous assessment of parallel structures.This approach alleviates the limitations of conventional parallel line detection methods that rely on overly simplistic constraints,and provides reliable structural primitives for subsequent SLAM feature construction.Second,to address the drift and failure issues commonly encountered by conventional point-line SLAM systems in static indoor environments with repetitive textures and cluttered object arrangements,a point-parallel line fusion SLAM method is proposed based on the detected parallel structures.By jointly constructing point and parallel line features and fully exploiting parallel constraints at both local and global levels,the proposed method improves localization accuracy and system stability.However,under extremely weak texture or complex illumination conditions,feature constraints may still be insufficient.To further enhance geometric representation,a multi-granularity SLAM framework that fuses points,parallel lines,and planes is introduced.Through planar parameterization and local Manhattan constraints,the method strengthens environmental geometric modeling while preserving the advantages of mid-level structural constraints.Experimental results demonstrate that incorporating parallel line features significantly improves localization accuracy and robustness across diverse indoor environments,particularly in visually cluttered scenes with partial regular geometric layouts.The proposed multi-granularity feature fusion further enhances mapping completeness and localization precision in complex static service scenarios.Third,to mitigate the adverse effects of moving objects on localization and mapping in real service environments,a dynamic-aware multi-granularity SLAM method is proposed by integrating a dynamic modeling mechanism.A feature-level dynamic probability model is constructed,in which dynamic attributes are assigned to multi-granularity features through observation likelihoods and temporal probability propagation.This enables reliable detection and removal of dynamic features,allowing the system to perform localization and mapping using only static multi-granularity geometric features.While preserving the benefits of rich geometric constraints,the proposed strategy substantially improves robustness in dynamic environments.Experimental results show that the dynamic feature filtering mechanism enables stable tracking in challenging scenarios such as crowded public spaces and open dining areas.Finally,a service robot experimental platform is developed for typical application scenarios,including residential environments,office spaces,and low-texture corridors.Extensive real-world experiments are conducted to validate the proposed system.The results demonstrate that the proposed multi-granularity geometric feature-based SLAM method consistently outperforms existing approaches in terms of localization accuracy,robustness,and environmental adaptability,highlighting its strong generalization capability and practical potential for real-world service robot applications.更多还原