基于多视图融合的闭环检测技术研究

Loop closure detection based on multi view fusion data

针对多视图融合点云片段的闭环检测问题,提出一种基于视觉词典的闭环检测方法,该方法避免了O(N2)的匹配复杂度的问题.首先对融合后的点云进行去除边缘响应等预处理.然后对每个点云片段提取尺度不变特征变换Scale Invariant Feature Transform (SIFT)关键点,计算快速点特征直方图Fast Point Feature Histogram (FPFH)描述子,将描述子空间离散化处理构建三维特征的视觉词典树,利用树状结构的词典加快了验证几何片段的对应关系.为了保证检测闭环候选系统的可靠性,采用了点云重叠区域作为几何验证的标准.最后,利用公开的数据集进行测试,得到了较高的召回率与准确率.实验结果证明了该方法可以实现自动的全局配准.

Aiming at solving loop closure detection of multi view fusion of point cloud fragments, a novel detection algo-rithm based on visual dictionary is proposed, which avoids a matching of complexity O（N^2）. Firstly, the fusion point cloud is used to remove the edge response and other preprocessing. Then the SFIT （Scale Invariant Feature Transform） key points and FPFH （Fast Point Feature Histogram） descriptors are extracted for each point cloud fragment. A 3D （three-di-mensional） visual vocabulary tree that discretizes a descriptor space is build. And use the tree to speed up correspon-dences for geometrical fragment verification. To ensure the reliability of detecting closed-loop candidates, the overlap ar-ea between the point clouds is used as the standard for geometric verification. Finally, experimental results on public 3D point cloud datasets demonstrate that the loop closure detection system has high recall rates and precision. The experi-mental results show that the proposed method can realize automatic global registration.