Research On Semantic Segmentation Of 3D Point Clouds In Indoor Scenes Based On Deep Learning

3D point cloud has been widely used in many fields such as autonomous driving and intelligent detection,and its efficient and accurate segmentation is the basis for good applications.Traditional segmentation methods need to manually filter features,and have shortcomings such as low efficiency and relying on manual prior knowledge,low segmentation accuracy,and inability to handle large-scale scenes.The development of deep learning networks provides technical support for solving the above problems.The current deep learning segmentation method pays too much attention to network construction,considers feature extraction from the global and local aspects,and does not pay attention to the repeated extraction of data category attributes and features;or converts point clouds into other forms(such as voxels).Convolution not only loses structural information but also imposes a great burden on computing resources;these practices lead to complex network structures but do not significantly improve segmentation accuracy.Therefore,starting from the nature of data and feature reuse,this paper designs a new lightweight semantic segmentation framework CSegNet,with the main line of improving the accuracy and accuracy of scene segmentation,and achieving high efficiency,intelligence and automation.The main research work and results are as follows:(1)Construction of global and local feature extraction modules.Aiming at the problem of insufficient utilization of local features of PointNet series,point-by-point convolution Point Conv is established as the backbone encoder to replace the MLP module in PointNet for feature extraction.The matrix operation obtains the local area features of the points,which enables the network to extract the global and local features of the point cloud more effectively.(2)The introduction of offset attention mechanism.In order to improve the problem of incorrect classification of objects,an operation of offsetting the point cloud is proposed,and the offset from the sampling point to its centroid is learned through the offset attention mechanism module,and then all coordinate points are converged and translated to the centroid according to the offset to obtain a new The category points are input to the next layer of network,so that the network can play a clustering effect during segmentation,thereby helping the network to better learn the characteristics of similar objects and improve the accuracy of semantic segmentation.(3)Reuse the edge convolution module.In order to improve the problem that the edges of objects in scene segmentation are not smooth,an edge convolution reuse method is proposed to capture edge feature information.By continuously updating the graph dynamically,the KNN algorithm is used for sampling points to find the K points with the nearest distance and calculate the edge features in turn.,and then use the shared network to update K features,and finally use the maximum pooling operation to integrate the K features into a new feature.It further strengthens the network’s extraction of point cloud edge features,and reuses the features effectively,which can smoothly segment the edges of objects.(4)Construction of the semantic segmentation network CSegNet.The network consists of two stages,each of which contains an encoding structure similar to PointNet++,but the point cloud feature extraction method uses point-by-point convolution.In the first stage,feature learning is performed to predict the centroid coordinate offset of each point,and then the offset attention mechanism is used to superimpose the origin cloud coordinate information in the second segment to segment each point in the point cloud;the second stage encodes part of the edge convolution module is added,and the decoding uses the deconvolution operation Point Deconv.(5)The performance of CSegNet network is verified by S3 DIS and ScanNet datasets.The results show that the segmentation of CSegNet on the S3 DIS dataset has achieved an overall accuracy of 88.7% and an average intersection ratio of 73.7%,which is not only accurate on large objects,but also in the face of small objects.The accuracy of PointNet++ is also improved by 9.8%;the average intersection ratio on the ScanNet dataset is 52.6%,and the segmentation accuracy of various elements is good.The superior performance on the two data shows that the CSegNet algorithm has stronger robustness and generalization ability,and the introduction of the offset attention mechanism and the edge feature module can utilize the point cloud features more effectively than the previous methods.