Research On LiDAR Point Cloud Semantic Segmentation Based On Dgcnn Network Model

In autonomous driving systems,semantic segmentation in Li DAR point clouds has grown in importance.As convolutional neural network technology advances,employing CNNs to semantically segment point clouds has also produced outstanding results.Li DAR sensors make use of laser ranging principles and are less impacted by lighting,item surface textures,ground clutter,etc.They also have high precision,long range,and high resolution.They have been widely used in research areas such intelligent robot navigation,positioning,obstacle avoidance,path planning,etc.because of their good stability and durability.They are a crucial component of intelligent vehicles’ environmental perception systems.Real-time semantic analysis of Li DAR point cloud data is a demanding research area because of the complexity of building structures and the dynamic diversity of roadways in urban contexts.Li DAR point cloud data contains features such as unstructured distribution,uneven density,and sparse distribution.As a result,this article has finished the following work to accomplish high-precision semantic segmentation of Li DAR point clouds:1)Combining bird’s eye view projections of polar coordinates with dynamic visuals.In addition to somewhat resolving the issue of more points being close to the Li DAR and fewer being far away,mapping the original Li DAR point cloud to the polar coordinate bird’s eye view image also largely preserves the spatial information of the Li DAR.Li DAR point cloud features may be extracted from dynamic images more thoroughly,which enhances network performance.2)Constructing a fresh feature map.The K Nearest Neighbor(KNN)algorithm was employed in the original technique to create the feature map during the dynamic picture extraction phase.To create the feature map in this article,a near neighbor point sampling and the Farthest Point Sampling(FPS)technique are combined.The network performance can be enhanced by include the FPS algorithm in the process of creating the feature map so that the network can capture both the internal and external contour features of the point cloud.3)The original Li DAR point cloud is projected into the polar coordinate bird’s eye view image and its aggregated features are learned using the dynamic graph network in the semantic segmentation network based on Dynamic Graph CNN(DGCNN).The initial Li DAR point cloud data is added to the aggregated features,which are subsequently sent into the convolutional network for additional computation.A significant number of repetitive calculations are required to accomplish high-precision Li DAR point cloud semantic segmentation,with an average intersection-over-union ratio as high as 56.5%.