The point cloud data of autonomous driving scenes collected by LiDAR is large in scale and contains rich spa-tial structural detail information,and some methods transform the point cloud to dense representations such as voxelization grids for processing,but ignore the information loss and occlusion problems caused by the point cloud transformations,which leads to degradation of segmentation performance.For this reason,this paper proposes a local feature aggregation net-works for 3D semantic segmentation.The local feature aggregation module therein aggregates the features of the K nearest points of the center point and obtains enhanced point features through a powerful attention mechanism,thus compensating for the lost information and improving the segmentation accuracy of the network.In addition,in order to improve the classi-fication accuracy of small objects,this paper proposes a 3D attention feature fusion block,which obtains richer semantic fea-tures and improves the performance of the network by discarding the conventional feature map splicing and using the atten-tion mechanism to decide the weights of different levels of semantic features.Extensive experiments on SemanticKITTI and nuScenes datasets demonstrate the superiority of the method.
国家科技期刊平台
NSTL
万方数据
