| With the concept of digital city and digital earth put forward,the processing and analysis of 3D data has become a research hotspot in the fields of photogrammetry,remote sensing,robotics and computer vision.As the basis of many applications such as scene modeling,autonomous navigation and object detection,visualization and semantic annotation of 3D scene point cloud are becoming increasingly important.In this paper,the cloud point data of campus scene obtained by Velodyne VLP-16 Li DAR(Light Detection and Ranging,Li DAR)sensor was taken as the research object.The semantic annotation method of campus scene based on conditional random field was studied,which includes the preprocess of scene point cloud,the point cloud segmentation and the semantic annotation of scene.The main work of this paper is summarized as follows:(1)Aiming at the problem that point cloud data obtained by Li DAR sensor contains a large number of outliers affecting subsequent classification,in this paper,through experimental comparison with the pass through filtering algorithm,the statistical filtering algorithm more suitable for campus scenes was used to preprocess the original point cloud data with 54,904 3D points collected.There were 6,767 3D points filtered,accounting for 12.33% of the original point cloud data,and the consuming time was 1535 ms.The outliers around the main objects such as buildings and trees were effectively removed,which achieves the purpose of rapid filtering.To solve the problem that the image contains a large number of useless ground points which increase the amount of subsequent computation,the random sampling consistency algorithm was adopted to segment the ground.Most of the ground points in the scene were quickly removed,a total of 11,821 ground points were removed,accounting for 24.56% of the number of 3D points before processing,which took 768 ms.(2)In order to reduce the number of computing nodes in the subsequent annotation model and improve the classification accuracy,maximum cliques need to be obtained through point cloud segmentation.In this paper,through experimental comparison with K-means segmentation algorithm,the region growth algorithm with simple principle and better segmentation boundary was select ed to segment the scenic point cloud,and 33 point cloud segmentation blocks were obtained,among which the plane points were more orderly,effectively avoiding over-segmentation.Then,using the obtained point cloud clusters,27 maximum cliques were generated through the maximum cliques construction algorithm based on regional growth,which took 192 ms.The number of nodes to be calculated in the annotation model was reduced from 35,137 single points to 27 maximum cliques,which laid a foundation for the improvement of classification efficiency.(3)In order to better realize the classification and annotation of main objects and display the results,in this paper a semantic annotation system for campus scene point cloud with four functional modules was designed and completed.The maximal cliques containing contextual semantic information was introduced to construct a conditional random field model,and the sub-gradient iteration method and graph cutting inference algorithm were used to learn and infer the parameters of this model.Then the six point features,one edge features and four maximal clique features were combined to realize the classification and annotaion of the scene point cloud,and the results of each step were visualized.A total of 146,579 3D points in the data set of 4 typical campus scenes were tested,and the overall classification accuracys were 95.06%,93.90%,94.12% and 93.74%,respectively.All of the overall classification accuracy were over 93%,and all of the annotation could be completed within 1500 ms.Finally,the random forest algorithm was used to conduct comparative experiment.The results show that the classification accuracy of this method proposed in this paper was higher,which meets the requirements of semantic annotation of campus scene point cloud. |
PDF Full Text Request