Semantic Understanding Method Of Autonomous Driving Scene Based On LiDAR Point Cloud

Autonomous vehicles are vehicles that can sense the environment and can drive without human intervention.With the rapid development and application of artificial intelligence,autonomous driving has gradually matured.LiDAR can obtain 3D spatial information of large-scale scenes steadily and quickly,and has become an important sensor for the understanding of autonomous driving scenarios.The LiDAR point cloud contains rich semantic information in the scene and is the main data type for understanding,analyzing,and interpreting autonomous driving scenarios.LiDAR point cloud has the characteristics of high density,high precision and mass.In recent years,research on semantic understanding of complex autonomous driving scenarios based on LiDAR point clouds has made considerable progress.Accurate and robust semantic understanding of autonomous driving scenes has important theoretical value and practical significance today.This article studies the semantic understanding of automatic driving scenarios based on LiDAR point clouds,and includes the following three aspects:1.In order to solve the current shortage of LiDAR point cloud data sets that contain a large number of complex scenes,this article first established the ScienceIsland data set.The data set contains a variety of complex scenes.The collection of vehicles passes bridges,buildings,and lakes along the way.There are a lot of vegetation,lawns,and trees on both sides of the road.There are many pedestrians,riders,and moving cars,trucks,etc.Meet the needs of semantic understanding of autonomous driving scenarios.Next,this paper proposes an accurate point cloud semantic segmentation framework.Four steps are achieved by converting the input point cloud into a distance image through spherical projection,semantic segmentation of a two-dimensional convolutional neural network,back-projecting the semantic label obtained by semantic segmentation of the distance image back to the original point cloud,and fine post-processing segmentation of the semantic edge Semantic segmentation of complete point clouds.The experimental evaluation of this paper shows that the improved two-dimensional deep convolutional neural network in this paper performs better than the current LiDAR point cloud semantic segmentation method on distance images,but it may lead to the problem of semantic edge blurring.To this end,in this paper,through semantic reconstruction of the original points,all laser scanning points in the original point cloud can get semantic labels.In addition,the efficient post-processing in this paper can also recover important semantic boundary information lost during projection.The method proposed in this paper takes a solid step towards the complete semantic segmentation of the LiDAR point cloud.2.A ground segmentation method based on convolutional neural network and probability-occupied grid is proposed,which can realize real-time,effective and accurate ground segmentation in complex driving scenarios.Aiming at the problem that the segmentation result is not accurate enough,the lightweight convolutional neural network has improved the rough segmentation effect of the ground points,and a good balance between accuracy and speed;at the same time,the cascading probability occupies the grid map modeling method,Use time series information to optimize the ground segmentation results;in order to improve the algorithm's robustness to complex environments,the method proposed in this paper divides the LiDAR raw data into three subsets according to its position:internal area,forward area and backward area;then,Iteratively reweighted linear square model uses estimated ground points to establish the slope function;finally,the ground segmentation of the forward and backward regions is realized based on the regional strategy to estimate the slope line.In view of the problem that the point-by-point segmentation results are too complicated in time and space and cannot be directly applied to the decision planning of autonomous vehicles,this paper uses an occupied grid map to represent and quantify the ground segmentation results.The experimental results show that the methods proposed in this paper are different.Complex scenes have good robustness and accuracy,and can realize ground segmentation in real time.3.A method for detecting negative obstacles based on point clouds in an off-road environment is proposed.The method includes the following steps:geometric feature extraction,local ground fitting,candidate point pair filtering,MSMF and candidate point clustering.The experiment was conducted under different terrain conditions.Experimental results show that the method has good robustness under different terrain conditions.The proposed method mainly has the following three contributions:A multi-LiDARs installation method for autonomous vehicles in an off-road environment is designed.Different from the traditional upright installation method,this installation method installs multiple LiDARs laterally on the roof of the vehicle to reduce the blind area around the vehicle body and increase the density of the scanning lines;in order to improve the robustness of the algorithm to detect negative obstacles in complex environments By estimating the vector of the ground scanning around the potential negative obstacle,adaptively adjusting the threshold,and combining the geometric characteristics and vector of the ground around the negative obstacle,the method proposed in this paper can detect the negative obstacle under different terrain conditions;This rule's iterative probability is also a contribution of this paper.In this paper,the vehicle pose is used to align multiple negative obstacle feature points on multiple LiDARs in time and space,and the negative obstacle features detected by multiple LiDARs in multiple frames are fused.The proposed method It can not only increase the maximum detection distance of negative obstacles,but also avoid the missed detection caused by the single LiDAR being blocked,avoid the missed detection caused by the blind area,and improve the detection accuracy.This topic is devoted to the research of semantic understanding technology of automatic driving scene based on LiDAR point cloud.A large number of actual vehicle tests have been carried out in the automatic driving scene of urban and off-road environment to verify the proposed algorithm's stickiness and accuracy.Finally,this paper summarizes the main contributions and innovations of this topic,and looks forward to the future research direction of semantic understanding of autonomous driving scenarios.