Research On Point Cloud Filtering Algorithm For Rain And Snow Weather Based On Vehicle LiDAR

In recent years,the rapid development of autonomous driving technology has driven advancements in sensor technology for autonomous driving systems.LiDAR,as a core sensor in these systems,can quickly and accurately capture large amounts of 3D information,providing essential perception information for autonomous driving.However,LiDAR’s detection performance can be negatively impacted by adverse weather conditions,such as rain and snow.Maintaining stable operation under adverse weather conditions is a significant challenge for autonomous driving technology to replace human driving.Rain and snow are typical adverse weather conditions that can affect LiDAR’s performance by introducing noise signals into the echo signal,which can interfere with the normal operation of the autonomous driving perception system,leading to safety hazards such as system misjudgment.Raindrops and snowflakes distributed in the detection range of LiDAR sensors can affect their detection performance,so LiDAR point clouds must be filtered and denoised before use.Conventional point cloud noise has corresponding filtering algorithms and good filtering effects.However,for noise in rainy and snowy scenes,these algorithms are difficult to achieve good noise removal effects and may mistakenly remove some environmental feature points,leading to the loss of environmental details.Rain and snow noise is dense at a closer distance to the LiDAR sensor,and the density of noise decreases gradually with the increase of distance.Intensity information is a measurement index used to measure the laser pulse echo intensity reflected by the target object detected by the LiDAR.Different target objects reflect different laser pulse echo intensities,which can be used as a feature information of the target object for recognition and classification purposes.However,classifying target objects based only on intensity information may lead to incorrect classification results,as target objects may have similar or even identical intensity information as other objects.Statistical filtering algorithms can achieve good filtering effects at close distances to the LiDAR sensor but usually lose environmental feature information at farther distances.To better solve the problem of rain and snow noise,this paper proposes a statistical filtering improvement algorithm that introduces intensity and distance information.Experiments show that the improved algorithm performs well in both rainy and snowy scenes by integrating intensity and distance information.The algorithm can further reduce the loss rate of environmental features while improving the noise removal rate and has good algorithm efficiency.The main contributions of this paper are:(1)We build a point cloud collection system to obtain LiDAR point cloud data under rainy and snowy weather conditions,manually label and classify point cloud data by each point using Cloud Compare point cloud processing tools,and use these data to form experimental and reference groups,providing data support for subsequent experimental verification work.(2)Based on the collected point cloud data,we perform equidistant segmentation on the data,analyze the spatial distribution characteristics of rain and snow noise points and the distribution characteristics of their intensity information by statistically counting the number of points and their intensity information in each distance interval range,and write an algorithm program to fit the gamma distribution mathematical function curve model to describe the spatial distribution law of rain and snow noise points.(3)We propose a point cloud filtering improvement algorithm for rainy and snowy weather scenes by integrating the advantages of existing point cloud filtering algorithms.By introducing intensity and distance information as constraint conditions in the statistical filtering algorithm,the improved algorithm can remove rain and snow noise while retaining more environmental feature points and consuming fewer computational resources,which has significant performance improvements in filtering accuracy and real-time performance compared to existing algorithms.(4)We evaluate the comprehensive filtering performance of existing point cloud filtering algorithms and improvement methods for point cloud data under rainy and snowy weather conditions,analyze the advantages and characteristics of the improved algorithm compared to existing algorithms.Experiments show that the improved algorithm achieves the best filtering accuracy and algorithm efficiency compared to existing algorithms in snowy scenes and also has good filtering effects compared to existing algorithms in rainy scenes.Furthermore,the improved algorithm exhibits higher robustness in practical application scenarios compared to existing filtering algorithms,and can effectively meet the requirements for point cloud filtering under adverse weather conditions.