Research On Laser Point Cloud Classification Of Outdoor Large Scene Based On Machine Learning

With the rapid development of laser scanning technology,dense 3D point cloud data can be obtained in outdoor large scenes.In recent years,the application of 3D point cloud data in the fields of autonomous driving,smart city and reverse engineering has also received extensive attention,and 3D point cloud classification is a key step in these fields.Therefore,it is of great significance to classify the point cloud under large outdoor scene.This paper mainly studies and improves the ground filtering algorithm in pre-processing point cloud,point-based point cloud classification algorithm and object-based point cloud classification algorithm.In the point cloud preprocessing stage,after establishing spatial index,removing noise points and downsampling the original point cloud,this paper proposes a ground filtering algorithm based on region growing.This algorithm combines principal component analysis to eliminate a large number of non-ground points.The ground filtering accuracy is improved,and the experiment proves that the ground filtering algorithm can ensure the integrity of non-ground points in the process of extracting the ground.Then,the point cloud classification based on points is studied.Firstly,the minimum entropy model is used to select the neighborhood range.Then,various features are extracted and merged in the support region,including the normal angle distribution histogram and the latitude direction sampling histogram proposed in this paper.After that,the features are trained and classified using a support vector machine,and the classification results are optimized according to the multi-scale neighbor tags.The experimental results show that the point-based point cloud classification framework proposed in this paper can effectively deal with the point cloud classification problem and at the same time,the prediction results have higher accuracy.Finally,the object-based point cloud classification algorithm is deeply studied.After using the DBSCAN algorithm to perform preliminary point cloud segmentation,the Kmeans clustering algorithm is used to further segment the point cloud objects,ensuring that the points in each object actually belong to the same category.After extracting the characteristics of each object,the SVM is used for preliminary classification.This paper proposes to use the initial classification information for MeanShift clustering,improve the clustering accuracy,and use the MeanShift clustering result as a high-order group of conditional random fields.Finally,it is classified by conditional random fields.Experiments show that using the object-based point cloud classification proposed in this paper,the results have higher classification accuracy.In the outdoor large scene,the point cloud classification algorithm of this paper separately studies the point-based and object-based methods.After experimental comparison,the object-based point cloud classification can obtain better overall classification results,but the effect in the adhesion area is still insufficient.It is more suitable for point cloud classification of outdoor scenes with larger object spacing.The point-based classification algorithm can achieve better results in details,and is more suitable for outdoor scenes where objects have certain contact.