Research On Ground Point Filtering Of Vehicle LiDAR Based On Gaussian Process Regression

Self-driving as the main direction of the current development of the automotive industry,it can provide important guarantees for future driving safety.Vehicle Li DAR is an indispensable technical in achieve unmanned driving,because it can actively obtain high-precision three-dimensional spatial information in large scenes.Ground point filtering is a key process for Vehicle Li DAR data processing,and it is also a necessary step to achieve unmanned driving with Vehicle Li DAR point cloud data.The Innovusion Li DAR used in this paper has characteristics such as uneven data distribution,inaccurate detection positions,missing data,and irregular data arrangement.It also analyzes common ground point filtering algorithms and concludes that common ground point filtering algorithms have high misrecognition rates,and this algorithm has a large dependence on the initial point cloud selection,and this algorithm has a poor filtering effect at the edge position.Therefore,a ground point filtering method for Vehicle Li DAR based on Gaussian process regression in this paper is proposed.This paper uses K-means method to optimize the filtering algorithm to solve the problem of poor applicability of the algorithm model in local area.The main research contents and conclusions of this article are as follows:(1)Point cloud preprocessing.This paper extracts one frame from the data for analysis,uses the least squares method and rotation matrix method to correct the data,and simplifies the data according to the angle of the laser scan.Experiments show that after point cloud preprocessing,the number of point clouds decreases,the point cloud regularity increases,the number of point cloud features increases,and the error angle between the actual position of the point cloud and the detected position becomes smaller.(2)Construction of Gaussian process regression model.By comparing the characteristics of different kernel functions,the square exponential kernel function is selected as the Gaussian model kernel function,and the DFP algorithm is used to optimize the hyperparameters of the Gaussian model.The experiments show that the ground point filtering algorithm based on Gaussian process regression is superior to other algorithms in the average misrecognition rate of point cloud and in actual filtering effect.(3)The K-means algorithm is used to perform spatial clustering on the preprocessed data.The algorithm predicts Gaussian models for each group of data,and uses the Bayesian weight allocation strategy to perform weighted fusion of the models to obtain a Gaussian process regression prediction model.The experiment proves that the model contains global point cloud features,which successfully improves the problem of poor edge filtering due to the concentration of training data distribution,and the algorithm works well in a variety of scenarios.