Self-supervised Learning For Semantic Segmentation Of Aerial Laser Point Clouds

Semantic segmentation of large-scale airborne laser scanning(ALS)point clouds is a basic problem for intelligent understanding of 3D scene.In recent years,great improvement has been achieved by training deep neural networks with supervised learning methods for point cloud semantic segmentation.Due to the spatial heterogeneity and category heterogeneity of ALS point clouds,the neural network suffers from model generalization disaster,that is,the deep neural network trained in one region is difficult to be directly used in other regions.Moreover,due to the extremely high cost of large-scale ALS point cloud semantic annotation,it is difficult to obtain high-quality labeled points covered wide areas.How to reduce the dependence of neural network on point cloud annotation and achieve high accuracy of ALS point cloud semantic segmentation has become the frontier issues of semantic understanding in the field of photogrammetry.This thesis studies the above needs and proposes an ALS point cloud semantic segmentation method based on self-supervised learning(SSL),which relieves model’s dependence for large-scale labeled points and learns the spatial information knowledge.The main research contents of this paper are concluded as follows:(1)This paper analyzes the characteristics of ALS point cloud,summarizes the results of deep learning network on the different ALS point cloud benchmark datasets and further results of semantic segmentation in different training subsets.Status of existing deep learning in ALS point cloud semantic segmentation are studied.These experimental conclusions provide guidance for the subsequent research.(2)This paper presents HAVANA(Hard neg Ati Ve s Amples Aware self-supervised co Ntrastive le Arning)model,a self-supervised contrastive learning framework for ALS point cloud semantic segmentation.Basic idea is to use the same location point clouds after data augmentation,which shares the same features and can be used as the data internal supervision signal to replace manual annotation,realizing the learning of general feature expression for semantic segmentation task.In terms of the selection of positive and negative samples in contrastive learning,this paper proposes the Abs PAN(Absolute Positive And Negative samples)strategy which uses k-means clustering to help the hardest negative sample selection.A large number of ALS un-labeled point clouds are used for representation learning,the underlying data information of ALS point clouds are mined to achieve the gain of the deep learning model in semantic segmentation tasks and reduce dependence on semantic labels.(3)A general point cloud dataset ALS-MLS for auxiliary tasks is constructed.The HAVANA has realized the improvement of ALS point cloud semantic segmentation.In order to further explore the potential of self-supervised learning,this paper re-construct dataset used in auxiliary tasks from the perspective of high-quality data.The study summaries the elements of high-quality dataset,unifies the scale and semantics of data from the multiple platforms.The ASampling(Active Sampling)strategy proposed for data selection as active learning.The ALS-MLS dataset achieves greater gain of self-supervised learning for semantic segmentation tasks.The experiments on two ALS benchmark datasets show that compared with the self-supervised learning Point Contrast,HAVANA method has an improvement of 3.4% in OA and 4.6% in F1.AVG on ISPRS Vaihingen data sets.Exceeding the benchmark supervised network KPConv at {10%,20%,40%,60%,100%} training subsets,especially when semantic labels are severely inadequate(10% of the training set),HAVANA can achieve more than 94% of the performance of the total training data.Ablation experiments confirmed the effectiveness of our self-supervised learning framework on multiple frameworks(Point Net++,Minkowski Net,KPConv).In addition to the LASDU dataset,our HAVANA method achieved state of the art in all five categories,demonstrating our selfsupervised method learning ability.To optimize auxiliary task datasets,this paper constructs a general dataset ALS-MLS.By comparing ablation studies of our construction strategy,the results demonstrate that our constructed dataset is superior to a single dataset.Finally,in many experiments,this paper surprisingly finds that the HAVANA has a significant effect on preserving geometry structure,and more importantly,such a method surpasses supervised learning methods in dense urban laser point clouds.