| With the rapid development of modern sensor technology,it is relatively easy to obtain high quality LiDAR point clouds.Nowadays,LiDAR point clouds are widely used in major 3D understanding tasks,such as classification,semantic segmentation and object detection;among them,semantic segmentation of LiDAR point clouds is an important step of high-level 3D point cloud understanding,and has great significance in the field of autonomous driving,robotics,Augmented Reality(AR)and smart city.At present,semantic segmentation methods of LiDAR point cloud can be divided into traditional methods and deep learning methods.Traditional segmentation methods mainly extract the features of artificial design,and then apply machine learning methods to model the features of design.However,handmade features rely on the prior knowledge of the designer and have poor generalization ability.Recently,deep learning-based methods have dominated the field of point cloud semantic segmentation,which can be divided into indirect methods and direct methods.Indirect methods transform point clouds to structural data and then apply convolutional neural network to segment.However,converting point cloud to these formats will lose the rich information of point cloud itself and many details,and will occupy a lot of memory.Direct methods can directly process original point cloud,which still contain a topic of learning effective point cloud features.Therefore,based on attention mechanism and multi-scale enhancement mechanism,this paper aims at solving the problem of effective feature learning of original point cloud based on deep learning,including effectively learning of both local features and global features.The contributions of this paper are following:(1)Based on existing searching,we propose a new multi-scale attentive LiDAR point cloud segmentation neural network.Upon a baseline encoder-decoder architecture which named Rand LA-Net,first,a point feature enhancement module is applied to enhance the inputting features.Second,an attentive skip connection is proposed to replace the commonly used concatenation to bridge the semantic gap between the point cloud features of encoder and decoder and ensure the global consistency of the same scale features of encoder and decoder.Third,a channel attentive enhancement module is proposed to boost the difference of local features and enhance local channel structure information.Four,our method develops a multi-scale feature aggregation method to the decoder to capture the global point cloud information of different scales.Finally,our method significantly outperformed effectively on the CSPC and Toronto3 D dataset.(2)We propose a new method which is named two-branch attentive aggregation network.First,we propose a new two-branch feature extraction frame,extracting global and local features meanwhile.Second,in the part of feature fusion of every branch,we introduce attention mechanism of transformer into this method to fuse information of selected areas.Finally,we test this method on DALES and Guangzhou datasets. |
