Point Cloud Multi-scale Contextual Relationship Metrics And Their Semantic Segmentation

Point cloud analysis has lately attracted increasing attention due to its wide applications in many areas,such as computer vision,robotics,and autonomous driving.Traditional point cloud analysis methods extract point cloud representation with artificial rules or hand-crafted features.However,the dependency on heuristic prior knowledge restricts their ability to handle real complex scenarios.With the rapid development of deep learning,many researchers focus on applying this technology to point cloud analysis and have achieved great success in many tasks.However,it is still a challenging task to efficiently extract the local features of the point cloud due to its natural irregularity,disorder,and sparsity.Meanwhile,point cloud semantic segmentation,as a fundamental vision task,plays an important role in practical applications.Existing semantic segmentation methods suffer from mis-segmentation and noise prediction in complex point cloud scenes due to the lack of sufficient semantic context information.To solve these problems,this thesis proposes two works to efficiently extract the local features of the point cloud and improve the semantic segmentation performance.Firstly,in order to extract local features of the point cloud more efficiently,a new network named Octant Convolutional Neural Network(Octant-CNN)is proposed,which consists of octant convolution module and sub-sampling module.For the input point cloud,the octant convolution module locates nearest points in eight octants of each point,and then transforms the geometric features into semantic features through a multi-layer convolution operation.The low-level geometric features are effectively fused with the high-level semantic features so that the local structure information can be efficiently extracted.The sub-subsampling module groups the original point set and aggregates the features to expand the receptive field of features,and also reduce the computation overhead the network.By stacking the octant convolution module and sub-sampling module,Octant-CNN obtains the feature representation of 3D point cloud from low-level to abstract,and from local to global.Extensive experiments demonstrate that Octant-CNN achieves great performance in four 3D scene understanding tasks including object classification,part segmentation,semantic segmentation,and object detection.Secondly,a Backward Attentive Fusing Network with Local Aggregation Classifier(BAF-LAC)is proposed to address the problem of lack of sufficient context information in point cloud semantic segmentation.It consists of a Backward Attentive Fusing Encoder-Decoder(BAF-ED)to learn semantic features and a Local Aggregation Classifier(LAC)to maintain the context-awareness of points.More specifically,BAF-ED bridges the semantic gap between encoder and decoder features.In each layer,encoder features are enhanced with an attention map derived from higher-layer features and then fused with the corresponding decoder features.LAC adaptively enhances the intermediate features in point-wise MLPs via aggregating the features of neighbor points into the center point.Equipped with these modules,BAF-LAC can extract discriminative semantic features and predict smoother results.Extensive experiments on Semantic3 D,Semantic KITTI,and S3 DIS demonstrate that the proposed method can achieve competitive results against the state-of-the-art methods.