The recent development of consumer-level RGB-D sensors increases the acquisition of indoor scan-ning data. However, the point clouds acquired by such low-quality equipment are usually noisy and incomplete. This paper presents an automatic algorithm for indoor scenes primitive abstraction and segmentation from sin-gle-view RGB-D scanning data. The first step is to preprocess the scanning data, which will automatically de-tect indoor structures such as walls, ceilings and floors. The input data is down-sampled and outliers are fil-tered out. The remaining point clouds are then abstracted in a semantic-aware manner and be represented by a set of geometric primitives. This abstraction step will produce the robust representation of indoor objects and parts, and reduce the computational cost of processing large-scale data efficiently. Finally, to yield the semantic indoor scene segmentation, a graph-based algorithm is adopted to group these primitives which incorporates feature descriptors and pair-wise geometric relations. Experimental results demonstrate that our proposed ap-proach can semantically segment the cluttered indoor scenes efficiently and robustly.%随着消费级RGB-D设备的普遍使用, 室内场景三维扫描数据更易获取, 但通过此类低分辨率设备获取的点云数据通常带有噪声且缺失严重. 为此, 基于单幅 RGB-D 扫描点云数据, 提出一种室内场景基元提取与自动分割方法. 首先对 RGB-D 扫描数据进行预处理, 自动检测场景中的墙面、天花板、地板等结构, 并对点云进行降采样和离群点滤波处理; 然后利用几何基元对剩余点云进行抽象, 通过几何基元的组合来鲁棒地表示室内物体和部件, 有效地减少大规模扫描数据处理的计算量; 最后根据每个基元的几何和颜色特征描述符以及基元之间的几何关系, 采用基于图的分割算法对基元进行组合实现室内场景物体的自动分割和提取. 实验结果表明, 该方法可以有效、鲁棒地抽象并分割杂乱的室内场景.
展开▼
