Semantic Segmentation Method Of Mls Point Clouds In Expressway Scenes Considering Feature Distinctiveness

Freeway is a very important traffic infrastructure,and accurate extraction and recognition of typical targets in freeway such as road surface,guardrail,vehicles,sign lines,etc.is of great significance in applications such as intelligent construction and management of freeway,automatic driving,and high precision map making.With the development of Li DAR hardware and point cloud data processing technology,semantic segmentation based on vehicle-mounted laser point cloud is gradually becoming one of the important technical means for highway target extraction and recognitionThe existing highway typical target extraction can be mainly divided into traditional artificially designed feature-based methods and deep learning-based methods.Among them,the former needs to design artificial features for different targets,which is not universally applicable.For example,the same class of facets may contain targets of different shapes,sizes and heights,such as sound barriers on both sides of the road,billboards on the outside of the road,traffic signs above the road,etc.To extract them,several different features and thresholds need to be set,which makes the extraction process tedious and inefficient,and cannot extract multiple targets at one time;while the deep learning-based method can be used by artificial neural networks to The deep learning-based method can learn the common distributed features in the data,which can realize the extraction of multiple targets end-to-end at one time.However,the farthest point sampling method commonly used in the current mainstream deep learning-based methods for point cloud downsampling is inefficient in large-field point cloud processing,while the random sampling method has high efficiency but suffers from the problems of key point loss and unstable coverage;meanwhile,the mainstream methods mainly target the geometric attributes of point clouds for encoding,and do not specifically target the utilization of additional attributes(e.g.,RGB)of point clouds.In the highway scene,the geometric attributes of the road surface and the sign line are similar,and it is difficult to distinguish them well by using geometric information only.To address the above problems,this paper constructs a deep neural network structure based on the point cloud uniform downsampling method of vehicle laser scan line from the data acquisition and storage methods of vehicle laser point clouds.The network can simultaneously ensure high sampling efficiency and stable spatial coverage.In order to overcome the misclassification problem of extremely similar geometric information such as sign lines and road surfaces,a local feature space coding structure is designed in the constructed network model in this paper,which can effectively enhance the differentiation of depth features.In addition,in order to better utilize the point cloud neighborhood information,this paper combines a local spatial coding module and an attention mechanism to enhance the ability of the network to retain geometric details.In this paper,a highway vehicle-mounted laser point cloud semantic segmentation dataset is actually collected and labeled to verify the effectiveness of this paper’s method.The experimental results show that the sampling method and local feature encoding space of this paper help to improve the effect of point cloud segmentation,and the effect of this paper is significantly better than the current mainstream methods: while maintaining the same spatial coverage,the time consumption of the uniform sampling method is about 4‰ of that of the farthest point sampling method;using m Io U and Io U as the extraction accuracy evaluation index,the method of this paper has a greater improvement in the category of marker lines and facets Io U is improved by 3.85% and 8.09% respectively compared with the optimal method in the comparison method,and the overall m Io U is improved by 4.39% for all categories.Finally,based on the proposed method,a prototype system for semantic segmentation of highway vehicle-mounted laser point clouds based on deep learning is developed in this paper.