Research And Application Of Efficient Point Cloud Semantic Segmentation

In recent years,the widespread use of 3D point cloud data in the fields of autonomous driving,robotics,remote sensing,and medical care has promoted the development of point cloud scene understanding and analysis technologies.3D point cloud semantic segmentation is an important supporting technology for point cloud scene understanding and analysis,which aims to assign the correct semantic category to each point in the point cloud.Recently,with the enrichment of point cloud data and the development of deep learning technology,many researchers have proposed 3D point cloud semantic segmentation network and showed good results,but there are still problems such as low computational efficiency and difficult deployment.Focusing on the aforementioned problems,this paper conducts research on efficient semantic segmentation methods for different point cloud data scales and application scenarios.The main contributions of this paper are as follows:(1)Research on small-scale point cloud semantic segmentation algorithm based on Transformer structure.For small-scale point cloud data,such as 3D shape and indoor scene,this paper proposes a Transformer model based on point-volume blending for the first time.First,aiming at the square computing complexity of the current point cloud transformer model,the first linear complexity attention variant for point cloud processing-patch attention is proposed.In addition,in order to fully exploit the advantages of multi-scale representation of point clouds,a lightweight multi-scale attention module based on voxels is proposed to establish the relationship between multi-scale features.Based on two attention modules,this paper propose the Patch Former network,which has achieved a speed increase of 9.2×compared with the current point cloud Transformer model on related datasets.(2)Research on small-scale point cloud semantic segmentation algorithm based on sparse attention mechanism.Although Patch Former has achieved high efficiency and excellent performance in a series of point cloud segmentation datasets.But there are still two drawbacks.One is that a lot of computing resources are wasted on empty voxel feature extraction.The second is that patch attention only pays attention to semantically similar points and does not pay attention to distant points,making it difficult for the model to capture the global shape information of the model.To this end,this paper proposes a CUDA-based sparse window attention module.This module establishes the mapping of feature indexes by establishing a rule matrix table so that the network can avoid the problem of empty voxel calculation.In addition,relative block-level attention is proposed to enhance the global information modeling ability by embedding relative position representations.Finally,this paper builds a PVT network model,and verifies the effectiveness and efficiency of the model on related datasets.(3)Research on large-scale laser point cloud semantic segmentation algorithm based on multi-representation fusion.For laser point clouds,most mainstream models use 3D sparse convolution to extract point cloud features,but this operator is computationally expensive and difficult to deploy.This paper proposes MRFNet,a fast,accurate,and easy-to-deploy radar semantic segmentation network.MRFNet is a multirepresentation fusion network,including distance image branch,polar coordinate bird’s-eye view branch and point branch.The point cloud is projected to a 2D grid using two image branches,and features are extracted using hardware-friendly 2D convolutions.To avoid the information loss caused by projection,the point branch is used to extract the original fine-grained features of points.Finally,a point-level fusion module is proposed to adaptively fuse image grid features with point features.Experiments have proved that MRFNet is effective and efficient,and the model has been deployed in an automatic driving system.