Research On Sparse Point Cloud Matching Method For Autonomous Driving Scenarios

With the continuous progress of autonomous driving technology,accurate perception and efficient processing of the surrounding environment become essential.Lidar is a widely used sensor in autonomous driving scenarios,and the point cloud data generated by it has rich spatial information.Due to the large amount and redundancy of point cloud data,directly processing raw point cloud data for positioning and matching in autonomous driving scenarios will bring a huge burden to the vehicle computing unit,which cannot meet the real-time requirements.In order to solve this technical problem,improve the efficiency of point cloud data processing in autonomous driving and ensure the quality of processed point cloud data.In this paper,we first study the sparse point cloud processing method,and then propose an improved sparse point cloud matching method.The main research contents of this paper are as follows:In this paper,a fusion ground filtering method is proposed to improve the traditional ground filtering methods,such as height filtering,normal vector filtering and RANSAC.These traditional methods have many problems: height filtering may delete some actual non-ground low altitude points by mistake;Normal vector filtering is easy to lead to misjudgment.RANSAC can deal with irregular terrain well,but it has high computational complexity and is sensitive to parameter selection.The fusion ground filtering method proposed in this paper integrates multiple techniques such as segmentation,morphological processing,dense matching and fusion strategies,which has higher comprehensiveness and accuracy compared with traditional methods.Specifically,the point cloud data is first divided into ground points and non-ground points by segmentation algorithm,and then the ground points are further processed by morphological processing method to extract ground feature information.Then,the point cloud data is mapped to a two-dimensional plane,and the dense matching method is used to accurately locate and fit the ground point cloud.Finally,the results of each stage were fused to achieve comprehensive and accurate ground filtering.Compared with traditional methods,the proposed method has the following advantages: firstly,it can deal with multiple ground types,such as flat ground,inclined ground and irregular terrain.Secondly,the method has higher accuracy,which can process and describe the ground point cloud more carefully and achieve higher accuracy.Finally,due to the phased strategy,the proposed method can process point cloud data efficiently and achieve faster ground filtering speed.In this paper,an Adaptive Density Farthest Point Sampling(ADFPS)method for point cloud sampling is proposed to preserve the overall structure information of the point cloud while performing sparse sampling.In the process of lidar scanning,the amount of original point cloud data obtained is usually quite huge,containing a large number of redundant points.If all points are used to construct a high-precision semantic map and extract point cloud features through deep learning models,it will lead to a huge computational burden.Therefore,it is necessary to preprocess the point cloud to generate sparse point cloud data frames for map construction.The key to generate sparse point cloud data is how to reduce the amount of point cloud data and retain the information of the original point cloud as much as possible.Therefore,this paper adopts the ADFPS method based on stratified sampling,local feature calculation and local feature weight adjustment.By calculating the curvature and density of the point cloud,the sampling density is adjusted to retain the key geometric features.The experimental results show that compared with the traditional sampling methods,the proposed ADFPS method can not only effectively retain the original point cloud information,but also significantly reduce the amount of point cloud data.In addition,the sampled point cloud is more evenly distributed,which helps to improve the accuracy of subsequent object detection.This paper proposes a dynamic object filtering method for point cloud.This method makes full use of the local neighborhood relationship of point cloud data,effectively excavates the structural characteristics of point cloud data,and realizes accurate filtering of dynamic objects.Due to the huge differences in shape and structure between dynamic objects,they are often difficult to be filtered by traditional point cloud processing methods.To this end,this paper introduces the attention mechanism into graph neural networks to adaptively adjust the weights between neighborhood points in order to capture dynamic objects of various shapes and structures.Through experimental verification on multiple datasets,the proposed method performs well in dealing with the dynamic object filtering task in point cloud data.Compared with traditional point cloud processing methods,the graph neural network combined with attention mechanism can deal with the diversity of dynamic objects more effectively and improve the accuracy and robustness of filtering.In addition,the proposed method also performs well in terms of computational complexity and running time,which provides an efficient and reliable solution for the task of dynamic object filtering from point clouds.In this paper,an improved Iterative Closest Point(ICP)matching method is proposed to solve the limitations of the original ICP algorithm.The improved method combines Fast Point Feature Histogram(FPFH)feature,Randomized Sampling with Accuracy Controls(RANSAC)method and dynamic threshold.The accuracy and robustness of point cloud registration are improved.FPFH feature is used to describe the geometric shape of point cloud and improve the accuracy of feature matching.The RANSAC method is used to estimate the transformation matrix and exclude false pairs of matching points.Dynamic thresholds are used to adaptively adjust the matching criteria to fit the point cloud density and distribution of different scenes.In this paper,the proposed methods are tested in detail on the public dataset KITTI and the data collected by the real car.In order to comprehensively evaluate the effect of the proposed methods,we compare these methods with the current mainstream point cloud processing methods.Experimental results show that the various methods proposed in this paper have different degrees of improvement compared with the current mainstream methods.