Lidar Point Cloud Semantic Segmentation

In recent years,optical detection technology has developed rapidly.Lidar has derived various scanning methods and has functions such as ranging,speed,and tracking on various platforms.Gradually,lidar has become the most important method for acquiring 3D point cloud data and a variety of scientific research topics based on lidar point cloud have become hot topics in the field of computer vision.Among them,outdoor point cloud semantic segmentation plays an important role in reconstructing three-dimensional maps of cities and assisting intelligent driving which attracts many scholars to participate in research.This paper has conducted in-depth research on outdoor point cloud semantic segmentation and outdoor point cloud semantic segmentation for cross-source migration.The main innovations are as follows:(1)Traditional semantic segmentation models often use a single-layer forward propagation method to transfer features.This forward propagation method does not fully utilize the information contained in the features of each layer.Besides,when there are too many layers,gradients may disappear.In response to this problem,this paper proposes the Point Bi-Connected Network.The output features of each layer will not only be passed to the next layer,but also to the next layer of next layer.In addition,the local geometric feature is designed based on the spatial neighbor information of each point in the point cloud.This feature is used as a priori features and will be concatenated with the original feature of the point cloud to be used for further neighborhood feature extraction.(2)The rapid development of lidars has also led to the diversification of lidar types.Due to the differences in scanning methods,detection methods and other factors,the point cloud data obtained by different lidars have a large difference in data distribution,resulting that direct model migration is often difficult to obtain satisfactory semantic segmentation results.This paper combines the idea of auto-encoder with the traditional semantic segmentation model,and proposes a point cloud semantic segmentation framework for cross-source migration.While using the semantic segmentation end to learn the semantic segmentation task,it also uses the auto-encoder to learn the underlying features of the data set to be migrated.This paper implements three different specific segmentation models for cross-souce migration.This paper validates our results on the Semantic3D data set and the NPM3D data set,having mIoU of 72.7%and 72.2%,respectively.In comparison with other algorithms,the point cloud semantic segmentation model for cross-source migratioin has improved the best migration effect of Semantic3D dataset to NPM3D dataset and the best migration effect of NPM3D dataset to Semantic3D dataset both by 4.4 percentage points,respectively.The research results of this paper can be used to semantically label the remaining data or another large-scale outdoor point cloud dataset after obtaining the semantic labels of some data in a large-scale outdoor point cloud dataset.