Research On Semantic Segmentation Of 3D Point Clouds Based On Enhanced Representation On Feature Space

As an important explicit representation of three-dimensional scenes,point cloud data can provide accurate spatial location information,which has been widely applied and promoted in many fields such as deep space exploration,earth observation,robot navigation,autonomous driving,and intelligent medical care.As a critical technology in scene understanding tasks,the point cloud semantic segmentation technology based on deep learning can effectively improve the service capability of current intelligent applications based on point cloud data and has important research significance.However,the current intelligent point cloud semantic segmentation algorithms have much room for improvement in terms of accuracy,model processing speed,and applicability to different scenes.To solve the above problems,this thesis analyzes the shortcomings of the existing point cloud semantic segmentation algorithms and conducts research from the perspectives of improving the edge accuracy of point cloud semantic segmentation,reducing the probability of missed segmentation and false segmentation,etc.By introducing self-supervised pre-training based on point cloud data,establishing a structured representation based on the geometric adaptation of the target,and adding spatial topological relationship learning,this thesis provides several feasible solutions for enhanced representation of point cloud features.The main research results and innovations achieved are as follows:(1)A point cloud semantic segmentation algorithm based on edge-aware pre-training is proposed.To address the problem of unsatisfactory class edge segmentation in point cloud semantic segmentation tasks,this work adds point cloud edge information as spatial a priori information to assist the learning of point cloud features from the data distribution characteristics of point clouds.At the same time,to address the problem of the lack of large-scale dense annotation information in point cloud data,this work implements self-supervised learning on large-scale unannotated point cloud data using a self-encoder in combination with a feature mask prediction method,and designs a pre-training task that can effectively enhance the edge-awareness of the model in conjunction with the point cloud semantic segmentation task.In addition,the work also develops an independent feature aggregation scheme to preserve the point cloud edge details by dividing the point cloud data in spatial dimension based on the point cloud edge information extraction,so as to generate a point cloud global feature map with more accurate boundary features for the point cloud semantic segmentation task.The experimental results show that compared with Point Bert,the best point cloud self-supervised pre-training learning method at present,the model method can achieve 1.2%and 2.2% m Io U improvement on Scan Net v2 and S3 DIS,respectively.(2)A point cloud semantic segmentation algorithm based on a structural-aware graph convolutional network is proposed.To address the problem of low fineness of feature learning in multi-category mixed regions in point cloud semantic segmentation tasks,this work starts from the basic local neighborhood division of point clouds and analyses the nearest neighbor search mechanism commonly used in current algorithmic models for direct processing of point clouds,and finds that this neighborhood search mechanism has the limitation of a relatively fixed range of neighborhood search within a single scale and low search efficiency.Therefore,this work proposes a weighted search strategy that takes into account local spatial information and feature similarity.At the same time,in order to explore the local features of point clouds with more differentiation,this work proposes a structure-aware graph attention mechanism based on a reasonable organization of point cloud signals using graph structure,and a feature analysis of the local spatial structure of point clouds from both the spectral and spatial domains,so as to achieve a fine-grained understanding of the features of complex regions of point cloud targets and better grasp the geometric structure of different point cloud targets.The proposed structure-aware mechanism enables a fine-grained understanding of the complex region features of point cloud targets,thus better capturing the geometric structure of different point cloud targets.The experimental results show that compared with KPConv,the best dynamic convolution kernel-based method at present,the proposed algorithm model can achieve 1.8% improvement in m Io U effect on the S3 DIS dataset.(3)A point cloud semantic segmentation algorithm based on the learning of spatial topological relations is proposed.Due to the irregular distribution and fragmentation of point cloud data,the problem of different appearance representations between the same semantic categories and similar local area representations between different semantic categories often arises when semantic segmentation of point clouds is performed.To address this problem of category confusion in the semantic segmentation of point clouds,this work implements weighted aggregation of local neighborhood features by constructing neighborhood confidence based on the learning and understanding of point cloud edge and local information,and innovatively proposes a geometric feature modulation mechanism to modify the learning structure of local features with low confidence to further improve the model’s ability to understand local regions of point clouds.At the same time,this work proposes a densely connected feature cascade network to learn the contextual relationship between different scales of point clouds,which can effectively preserve the learning results of high and lowdimensional features while building long-range information dependency to reasonably diminish the misinformation propagation between different categories,in order to address the shortcomings of current hierarchical networks that only use the last scale of features to build global features.Experimental results show that the algorithmic network model proposed in this work can effectively improve the situation of segmentation errors between similar semantic targets in point cloud semantic segmentation tasks.On the point cloud part segmentation Part Net and Shape Net Part dataset,compared with the current best performance Closer3 D and KPConv algorithm,the improvement is 0.7% and 1.2% respectively.On the large-scale indoor point cloud scene segmentation dataset S3 DIS,compared with the current algorithm model Point Transformer with the best performance,the overall accuracy can be improved by 0.5%.At the same time,this method also effectively enhances the processing capacity of the algorithm model for the input of point clouds in different working conditions,and shows good robustness.In summary,the thesis proposes innovative and improved methods for the enhancement of initial feature descriptions of point clouds,dynamic local neighborhood segmentation,and learning of spatial contextual relationships between local regions from the perspective of feature space augmentation representation through an in-depth exploration of the main workflow of the point cloud semantic segmentation task.Experimental validation on publicly available datasets confirmed that the work described in this thesis can effectively improve the problem of inaccurate edge segmentation,segmentation errors in complex regions,and category confusion in the semantic segmentation of point clouds,providing a feasible solution to further achieve accurate 3D scene analysis and understanding.