| The depth information contained in 3D images can accurately restore the spatial structure of indoor scenes,and with the development of depth sensing technology,it has been widely used in many fields such as medical diagnosis,intelligent manufacturing and industrial production.Instance segmentation,as an advanced computer vision task that combines both target detection and semantic segmentation,has great potential for application.Although some research results have been achieved by existing methods,the instance segmentation results returned by the models still need to be improved in terms of granularity.In this paper,we conduct research on instance segmentation methods for 3D point clouds and design and implement a point cloud instance segmentation algorithm for use in indoor scenes,including the following:(1)To address the problem that conventional 3D convolution generates huge amount of operations and dilutes semantic information,this paper proposes the codec network incorporating subflow-type sparse convolution and double attention mechanism,based on which a point cloud semantic segmentation method in indoor scenes is implemented.The main body of the method adopts the U-shaped codec framework,and based on the fusion of substream sparse convolution,the dual-attention module is designed to further enhance the network’s ability to capture information at a distance,which uses parallel computation of channel attention and spatial attention to recalibrate the feature map.The module is also embedded in each layer of the encoder and decoder to form the new network architecture,and comparative experiments are conducted on the benchmark dataset Scannet(v2)for indoor scenes.Compared with the classical network Minkowski,the m Io U of this paper’s method is improved by 3.2%,which serves as an effective guide for the next step of implementing point cloud instance segmentation.(2)To address the problem of how to learn effective instance features from point cloud data with semantic category labels,the point cloud instance segmentation method based on semantic segmentation is proposed in this paper.The method takes the category labels output from the point cloud semantic segmentation network as input,and first designs the semantic auxiliary module in order to pass the labels and correct the instance centres.This module returns semantic category labels and instance centre bias vectors through two parallel two-layer perceptrons to pass semantic labels and correct instance centre vectors;then the hierarchical cluster learning module is designed to generate preliminary instance segmentation suggestions through high-dimensional feature space constraints;and finally the cyclic optimisation module is designed to optimise the instance segmentation results through cluster computation.By fusing these three modules,the semantic-assisted point cloud-based instance segmentation network is constructed and analyzed in the comparative experiment on Scannet(v2).Compared with the classical network SSTNet,the m AP of the method in this paper is improved by 3.3%and the average running time is reduced by 59 ms,which effectively improves the accuracy and efficiency of point cloud instance segmentation in indoor scenes.(3)In order to verify the effectiveness of the method proposed in this paper,the 3D segmentation system for indoor scenes was designed and implemented using the Python web framework Django.The system mainly includes basic information management,model parameter management,point cloud semantic segmentation module and point cloud instance segmentation module,and the interface and visualization prediction effect of each module are demonstrated.In summary,this paper proposes the point cloud semantic segmentation method based on dual attention coding and decoding networks and the point cloud instance segmentation method based on semantic assistance,which has certain theoretical significance and practical application value,in view of the problems of current 3D segmentation methods for indoor scenes. |
PDF Full Text Request