| With the development of autonomous driving technology,the research of laser radar sensor technology is becoming hotter.Lidar is also called optical radar,which is the abbreviation of laser detection and ranging system.It determines the propagation distance between the transmitter and the target object,and analyzes the reflection energy size of the target object surface,the amplitude frequency and phase information of the reflection spectrum,so as to show the precise three-dimensional structural information of the target.It has the advantages of high resolution,strong anti-interference ability,and wide range of detection.The data format of Lidar is mainly point cloud,and the research of point cloud occupies an important position in the intelligent driving perception system.It is an important part of the real scene understanding.Deep learning is one of the most important breakthroughs in artificial intelligence research in recent years and has achieved great success in the fields of two-dimensional images(object classification,object detection,object segmentation,and object tracking).In the initial stage,the point cloud contains rich information such as threedimensional space coordinates,color,intensity,etc.However,due to its own characteristics of the disorder,irregularity,and uneven distribution,it is impossible to use the traditional deep convolutional network to directly act on the point cloud to extract features.Therefore,this paper proposes a new multi-layer perceptron(MLP)based local feature fusion network LAF-Net,which consists of a random sampling module,a local self-attention module,and a local-global feature fusion module.LAFNet samples the point cloud through a random sampling strategy to reduce the number of point clouds,which can greatly reduce noise and retain the overall information of the point cloud.The self-attention module can well extract low-level texture informations of the point cloud and pay more attention to the topological structure between the local points.The local-global feature fusion module can well gather each point and the adjacent points as well as the far end.The interdependent information between the contexts,the multi-resolution and scale features of the point cloud can be obtained through the multi-level accumulation of the module.On this basis,due to the large number of parameters of the method based on the multi-layer perceptron,and the random sampling will lose a large amount of information of the key points of the point cloud,the local information aggregation ability of the point cloud is insufficient,so a new point cloud is proposed.Cloud multi-head attention mechanism Transformer lightweight network Pointmer,the network has linear complexity,while reducing computational complexity and reducing the number of parameters while maintaining performance benchmarks,this study shows on four different public benchmark datasets,LAF-Net network and Pointmer network can effectively extract the geometric semantic information of point cloud,and perform well in point cloud classification,segmentation and other tasks.The innovations of this study are as follows:(1)A new network of local features based on multi-layer perceptual(MLP)LAFNet,which is very suitable for research on point cloud learning.(2)The LAF module uses low-dimensional feature branches to obtain low-level topology information and high-dimensional feature stitching after position coding,so that low-level texture features and high-level features can interact to obtain multi-scale and multi-resolution semantic fusion information.(3)A new light network for point cloud based on Transformer have been proposed,called Pointmer.The network has linear complexity O(N),which can reduce the computational complexity and reduce the number of parameters while keeping the performance benchmark unchanged.This network can be used as the basic backbone network of point clouds to extract relevant features and adapt to the learning and processing of point cloud.(4)Experiments have proved that our networks have achieved excellent performance on S3 DIS,Modelnet40,Semantic3 D public datasets. |
