Research On Semantic Labeling Of High-resolution Laser Scanning Point Clouds

Nowadays,integrating laser scanners,position and orientation systems,and optical camera systems,modern mobile laser scanning(MLS)systems collect not only large-scale high-resolution 3D point clouds with real-world coordinates and high precision,but also a set of optical referenced images calibrated with 3D point clouds.In contrast to 3D point clouds with low-resolution,3D point clouds with high-resolution provide more detailed 3D geometry and shape information of objects in the scenes.Therefore,the researches on high-resolution MLS point clouds attract wide attention.Semantic labeling aims to assign accurate category labels to objects in the scenes,and accomplishes object recognition and segmentation.Automatically and effectively implementing semantic labeling is an essential approach for information extraction,and is required urgently in various areas,including urban planning,autonomous driving,map navigations,digital cities,etc.Therefore,this dissertation mainly focuses on the researches of semantic labeling based on high-resolution laser scanning point clouds.However,there are lots of challenges including huge amount of points brought by high-resolution point clouds,difficulties in manually creating training data sets from 3D point clouds,complex scenarios existing in the real world,etc.To handle these aforementioned challenges,this doctoral dissertation mainly focuses on the following three aspects:Firstly,to handle complex scenarios in 3D scenes and relieve computational burden caused by huge amount of points,a 3D patch-based semantic labeling framework is proposed.To transfer category labels between complex scenes efficiently,a 3D patch-based match graph(3D-PMG)structure is proposed to cluster 3D patches for alleviating the misclassifications caused by local similarities between different categories.To further consider contextual information among the spatially neighboring 3D patches,the proposed method exploits a set of stochastic label transfer strategies and contextual consistency constraints to achieve semantic labeling in complex scenarios robustly and efficiently.Secondly,to alleviate difficulties in obtaining training data sets of 3D point clouds,an active learning(AL)framework is proposed to iteratively select a small portion of unlabeled points to query their labels,and creates a minimal manually-annotated training set.To handle the biased sampling problem caused by category imbalance and local similarities,a neighbor-consistency prior is used to conduct unbiased sampling for selecting the valuable samples into the training set.Additionally,to reduce the number of categories used in labeling,a higher-order MRF containing a regional label cost terms,is exploited to refine the labeling results.Thirdly,in order to exploit the annotated images for achieving point clouds labeling effectively,a cross-dimensional label transfer method is proposed.To transfer the annotations from 2D image databases to the 3D point clouds effectively,the proposed method designs a probabilistic graphical model to encode the consistency constraints between multi-view images and 3D point clouds.To obtain a high performance of labeling 3D objects,the proposed method exploits convolutional neural network to detect objects on multi-vew images,and models geometry consisitency in 3D point clouds,textural consistency in 2D images,and spatial consistency between 2D images and 3D point clouds.The quantitative and qualitative evaluations of the three proposed methods are implemented on large-scale high-resolution laser scanning point clouds acquired from real world.The experimental results demonstrate that our proposed methods can conduct semantic labeling of high-resolution point clouds effectively.In addition,the comparative experiments also exhibit the superior performance of our proposed methods on labeling high-resolution point clouds.Therefore,our researches in this dissertation assist in further development of semantic labeling on high-resolution laser scanning point clouds.