This paper focuses on noisy points filtering in dynamic reconstruction. RGBD cameras can acquire depth information directly, which is an effective way to obtain a global point cloud. However, the inevitable error leads to erroneous rebuilding, which is mainly caused by wrong depth information and dynamic disturbance. Different from algebraic methods such as median filter, a geometric method is proposed based on common-view geometry. Noisy points generate more intersections with other projection lights, which are used to distinguish incorrect points. Considering the huge amount of computation, a sampling method is adopted to avoid unnecessary calculations and a cuboid method to simplify the calculation. We also employ a hybrid strategy with two different standards to increase the robustness when filtering. By calculating the geometry relationship effectively and robustly, we get good results in removing map points with incorrectly measured depth and caused by dynamic disturbance. Compared with the algebraic methods, the proposed method can also remove dynamic objects. The performance is verified in three popular public datasets.

本文着重于动态重构中的噪声点滤波。RGBD摄像机可以直接获取深度信息，这是获取全局点云的有效方法。但是，不可避免的错误会导致错误的重建，这主要是由错误的深度信息和动态干扰引起的。与中值滤波等代数方法不同，提出了一种基于共视几何的几何方法。噪点会与其他投影光产生更多的交点，这些交点用于区分不正确的点。考虑到大量的计算，采用采样方法以避免不必要的计算，而采用长方体方法来简化计算。我们还采用具有两种不同标准的混合策略来提高过滤时的鲁棒性。通过有效而稳健地计算几何关系，我们在去除深度测量错误且由动态干扰引起的地图点时取得了良好的效果。与代数方法相比，该方法还可以去除动态对象。在三个流行的公共数据集中验证了该性能。