Biography
Enrollment Date: 2013
Graduation Date：2016
Degree：M.S.
Defense Date：2016.05.30
Advisors：Xiang Xie
Department：Institute of Microelectronics，Tsinghua University
Title of Dissertation/Thesis：Segmentation method for point cloud of large-scale scenes based on measurement capability
Abstract:
Point cloud segmentation is the necessary premise for 3D reconstruction of large-scale scenes. Large-scale scenes often contain all kinds of objects and the spatial relationship between different objects is very complex. Moreover, the amount of point cloud data is extremely huge. Therefore it is still challenging for those existing methods to segment point cloud of large-scale scenes with both high speed and high accuracy. Through analyzing the working principle of Point Cloud Acquisition (PCA) system, we found that the spatial relationship of neighbor points on the same object is not only decided by the shape of the object, but also decided by the measurement capability of PCA system.Therefore, the measurement model of the PCA system is first given out in this paper, and the relationship between spacing distance of neighbor points and measurement capability of PCA system is analyzed in detail. Based on this analysis, a novel adaptive segmentation method fully utilizing the measurement capability of PCA system is proposed for large-scale scenes with high speed and accuracy.The proposed segmentation method mainly includes the following innovative key algorithms: 1) Ground filter algorithm: based on a new scan-line-based ground filter algorithm which uses the elevation, gradient information and the spatial relationship of neighbor points on the same object, ground points are accurately filtered with high speed and complex ground conditions such as slopes and bumpy roads can be well handled. Experiment results of different scenes show that the processing speed of the ground filter algorithm is over 1 million points/sec and the accuracy is 99.3%.2) Non-ground segmentation algorithm: to further segment remaining non-ground points fast and accurately, a novel measurement-capability-based self-adaptive segmentation algorithm is proposed based on the segmentation inequations. Different objects of complex large-scale scenes can be correctly segmented and the algorithm is of high efficiency by using the inherent ordering of point cloud. Experiments show that the average processing speed of the non-ground segmentation algorithm is 10.6μs per point and there is almost no under-segmentation happen.3) Merging algorithms for over-segmented clusters: After the non-ground segmentation, over-segmented clusters including over-segmented non-floating clusters and floating clusters need be further processed. A volume-based adaptive merging algorithm is proposed to merge over-segmented non-floating clusters; a minimum spanning forest (MSF) based merging algorithm is proposed to merge over-segmented floating clusters.Experiments show that the average precision and recall of the proposed method are 95.7% and 96.5% respectively, and the average processing speed is 45.8μs per point. All these performance indexes are overall superior to the popular segmentation methods.