Research On Point Cloud Segmentation Technology Of Road Environment Based On Deep Neural Network

With the popularization of artificial intelligence and the breakthrough of sensing technology,unmanned driving technology has developed rapidly in recent years.As a major sensor for unmanned vehicles,LiDAR is very important for understanding the road environment.Aiming to ameliorate the problems of disorder,sparseness and floating occur for 3D LiDAR point cloud in the road environment,we design a semantic segmentation algorithm based on deep neural network,also realize the detection of road drivable areas and the identification of dynamic obstacles accurately.Our main research results and innovations are as follows:In a spherical coordinate system in which the position of the LiDAR is the center of the sphere,the index coding of the point cloud 3D position in both vertical and horizontal directions is performed,which convert a 3D point cloud into a multi-channel point feature map.Through the campus road detection experiment based on Fully Convolutional Network and the point cloud classification experiment based on Point Network,we analyze the correlation between multi-channel feature maps in road scene and the two sensor data: color image and model point cloud,providing theoretical basis for point cloud semantic segmentation network to be designed.Aiming at the spatial correlation and feature richness of point cloud multi-channel feature maps,we designed a point cloud semantic segmentation network for road environment awareness: the design has special convolutional neural network such as jump connections and point micro-perceptron,which can solve the problem of empty hole and detail loss in the feature map.We design a convolutional network for basic feature extraction and a deconvolution network for multi-layer semantic feature fusion,and conduct the detailed experiments for ablation study,prove the validity and necessity of related structural design.Based on this,a high-real-time and high-resolution variant model is designed to meet different mission requirements.The experiments are divided into offline simulation test and physical platform measurement.We compare multiple mainstream methods on public data sets,the results show that the proposed point cloud semantic segmentation model has better detection performance,improving the accuracy and robustness for detection while taking care of realtime.We design a point cloud sensing platform based on ROS robot system combined with point cloud analysis,running algorithm for recognition and 3D visualization in road scene,and conduct test on unmanned golf carts,analyze the effect of detection in multiple typical road scenes,prove the applicability of the system designed in this paper to the perception of the road environment.