Research On Point Cloud Segmentation Method Based On Multi-Scale Feature Fusion

With the rapid development of three-dimensional data acquisition equipment such as lidar and depth sensor,the cost of obtaining three-dimensional point cloud data is gradually decreasing.Compared with two-dimensional images,three-dimensional point clouds can retain rich and real geometric features on the surface of objects,so they have gained extensive attention in the field of computer vision.3D point cloud information processing mainly includes 3D point cloud reconstruction,3D point cloud classification,3D point cloud denoising and 3D point cloud semantic segmentation.Among them,point cloud semantic segmentation is the core of point cloud information processing technology,and its purpose is to classify each point in three-dimensional point cloud data so that computers can understand the geometric shape,structure and semantic information of objects in point cloud.At present,three-dimensional point cloud semantic segmentation algorithms can be divided into two categories,namely,traditional methods based on manual design features and methods based on deep learning.Traditional point cloud semantic segmentation methods mainly rely on geometric information constraints,similarity attribute aggregation and statistical methods to divide non-overlapping segmentation areas,and then manually classify them on this basis,and finally get the semantic labels of the segmentation areas.However,this kind of traditional point cloud semantic segmentation method is generally limited by the problems of uneven density,noise sensitivity and poor robustness of point cloud data sets.Although these methods are simple and efficient,their segmentation results are usually not ideal.Compared with the traditional methods,the point cloud semantic segmentation method based on deep learning extracts and learns the features of enough points in the data set,and trains a neural network model with generalization characteristics to predict the semantic tags of points.The semantic segmentation method of point cloud based on deep learning can be divided into indirect processing method and direct processing method.The method of indirectly processing point cloud needs to regularize point cloud,which will cause a lot of feature loss.The method of directly processing point clouds needs a fine feature extraction framework to extract the geometric features contained in point clouds.However,this method is easy to lose a lot of point cloud details in exploring the correlation between local and global features,the fusion of advanced abstract features and primary semantic features on multiple scales,and the design of network structure.Therefore,it is still a challenging task to design a finer network to complete the semantic segmentation of point clouds in complex scenes.In order to solve the above problems,this paper proposes a 3D point cloud segmentation method based on deep learning by using attention mechanism and global feature relation matrix,that is,a point cloud segmentation method based on dynamic attention and global feature selection for multi-scale feature fusion(DGM).The feature extraction module of DGM consists of dynamic attention selection unit and point-based feature relation matrix aggregation unit.Dynamic attention selection unit uses dynamic graph and improved attention mechanism to adaptively extract and fuse the features of points in the neighborhood,so as to better capture the local features of point cloud data.However,the point-based feature relation matrix aggregation unit generates a global feature relation matrix by distinguishing the differences and connections between geometric dimension aggregation features and feature dimension aggregation features,so as to further adjust the features extracted from the dynamic attention selection unit,and finally integrate geometric features and independent features of points to further enrich the semantic representation of points.DGM combines three multi-scale features: geometric dimension and feature dimension aggregation feature,local aggregation feature and global relationship feature,abstract advanced feature and primary semantic feature,so as to enhance the correlation between various feature representations,thus improving the network’s ability to identify point cloud categories.On the whole network architecture,DGM adopts encoder-decoder structure and assigns high-level semantic features to each input point.DGM can effectively solve the problem of multi-scale feature fusion,improve the feature extraction ability of 3D point cloud data,and thus improve the accuracy of point cloud semantic segmentation.The main work of this paper includes:(1)In order to expand the receptive field of the center point in each neighborhood and strengthen the connection between neighborhoods in each layer of network,this paper uses dynamic graph to sample,so that the feature aggregation process of local areas is no longer limited to a fixed three-dimensional geometric distance.(2)In order to distinguish the features between neighborhoods more accurately and make the updated local aggregation features more representative,this paper adjusts the attention coefficient across the feature space of different points based on the existing attention mechanism,and improves the aggregation mode of attention mechanism.(3)In order to reduce the deviation caused by sampling and aggregation only depending on the difference of feature distance,this paper proposes a novel point-based feature relation matrix aggregation unit.It combines the aggregation features based on fixed geometric distance sampling and the attention aggregation features based on dynamic feature distance sampling to generate a global relationship matrix.The matrix can not only be used as an indicator for further adjustment of dynamic attention aggregation features,but also contains long-distance semantic information,which can enhance the model’s ability to distinguish complex point cloud scenes.(4)In order to avoid feature loss and improve the semantic representation ability of features,residual feature fusion is introduced at the end of feature extraction module.Specifically,it takes the independent features of each center point and the geometric features of local areas as residual features and fuses them with the previously extracted features.This method can improve the segmentation accuracy of part segmentation and local scene segmentation at the boundary.In this paper,experiments are carried out on Model40,ShapeNet and S3DIS data sets to verify the performance of DGM in point cloud model classification,object parts segmentation and semantic segmentation of large indoor scenes.Experimental results show that the proposed algorithm can achieve or exceed the performance of existing advanced methods.