Robust Method for Static 3D Point Cloud Map Building using Multi-View Images with Multi-Resolution

Abstract

Robots can perform various missions in multiple changing environments. The dynamic objects have significant influence on the long-term autonomy and 3D map construction, because “ghost tracks” inevitably exist due to the continuous error-accumulation of the input data. So it is critical to keep only static subsets and exclude noisy obstacles to mitigate the influence on mapping and navigation. We propose a robust static map building method, which compares the discrepancies between single scan data against the noisy map. This method focuses on the advantages of most dynamic objects of different views with unique attribution and will be easily detected in these views. Accordingly, we present the novel “Multi-View and Multi-Resolution” image-based method with BEV-RV (Bird's Eye View-Range View) modules to discriminate static/dynamic point clouds. Through two stages of iteration with different image window sizes of point level, we first collect more static points of some inevitably wrong judgments and then remove such completely unreliable dynamic points at a later stage. Experimental evaluations are conducted by using the KITTI dataset as ground truth. Qualitative analysis indicates that the proposed method is robust and reliable against state-of-the-art methodsin some dynamic regions.