Contextual Classification For Mobile LiDAR Point Cloud And The Detection Of Individual Objects

The classification of point cloud data and the extraction of individual objects in point cloud scenes is very important for automated 3D scene analysis,which is of great significance in the areas of 3D digital city reconstruction,urban planning,traffic signs detection,land use change detection and autonomous driving.High-accuracy point cloud data of complex scenes can be easily acquired with the improvement of the performance of 3D laser scanning system,especially for Mobile Laser Scanning system,which makes the point cloud widely used in automatic analysis tasks of complex urban scenes.It is urgent to develop an automated point cloud information extraction system,in which the point cloud classification is the key step.However,the special data behavior and data distribution of point cloud data bring great challenges to point cloud classification.The main difficulties are follows:(1)The extraction of point cloud features relies on point's neighborhood,and the point cloud that obtained by threedimensional laser scanning is a discrete points set,which does not have a regular structure,and the distribution of point cloud density in the scene is uneven,so that the size of the neighborhood is easily affected by the distribution of point clouds.Although the existing adaptive neighborhood estimation method has been able to restore a reasonable neighborhood of point cloud data,the insensitivity of adaptive neighborhoods still exist,especially in high-density regions where the neighborhood is linearly distributed.This may cases noise problem in the classification results.(2)The point cloud data is a digital reproduction of the natural scene,and there are abundant geometric features for point cloud classification task.However,the types of objects in the natural scene are complex and diverse,and the data missing problems caused by occlusion,both of them can easily result in incomplete classification results and prone to misclassification.Its indeed to take higher level statistics information of point cloud as a constraint in classification task,which cloud result an accurate and complete classification results.Aiming at the above key issues,researches have been carried out from two aspects in this paper: adaptive neighborhood estimation and multi-scale contextual classification.The adaptive neighborhood focuses on getting a reasonable neighborhood with the constrain of data distribution.The multi-scale contextual classification smooth the individual classification result by high level information in point cloud scene.Therefore,a relatively complete and feasible technical solution of point cloud classification can be obtained.Based on the point cloud classification results,the further work of the individual objects extraction is designed.The main work of this article can be concluded as follows:(1)Research on curvature based adaptive neighborhood estimation method.Classical adaptive neighborhood estimation method recovering neighborhood by traversing a fixed interval and select a best one.Since the distribution of point cloud density is not considered,the neighborhood in the high-density region is likely to present a linear behavior that the points within neighborhood arranged in a line.And the neighborhood of sparse regions may have a large size,which makes the feature unstable and has a significant impact on the classification results.This paper assumes that point cloud data can be roughly divided into irregular regions with large curvature and regular regions with small curvature,then a curvature based adaptive neighborhood estimation method is proposed.The method first divides the point cloud data by the curvature threshold.When the curvature of the point is larger than the threshold,the point is classified as an irregular point,and the curvature is smaller than the threshold is classified into regular region.Then a divide and rule based strategy is used to derive neighborhood results.This strategy choose different neighborhood estimation methods for different region,and corresponding parameter intervals are set with the constrain of point cloud density distribution,then the final adaptive neighborhood can be achieved.In this paper,a random forest classification method is used for the classification task to verify the effectiveness of the proposed curvature based adaptive neighborhood.Experiment results on public standard data sets proved that the proposed adaptive neighborhood estimation method is quite effective.The average evaluation values is more than 3% higher than the two classical methods.(2)Research on multi-scale contextual point cloud classification method.The classification method based on geometric features of point cloud data is easily affected by data missing and uneven distribution of point cloud.Except the geometric features,there still have a lot of high-level information in point cloud scenes.This type of information can be used as an constrain for the optimization of individual classification result,then a smoother and complete classification result can be obtained.This paper first improves the classic dependency context with an improved edge structure based enhanced local dependency.The classical method uses the k-nearest neighbor structure as the edge structure in the global region.While the proposed method constructs an improved edge structure in the regular region.This edge structure insure the neighboring structure span at least two scan lines in the regular region to capture long-distance dependencies in the scene.Meanwhile,a super-voxel entity based distributed spatial contextual statistical method is proposed,which use a property statistics strategy and stacking methods to simulate the horizontal and vertical distribution of objects in the scene.Then,the multiscale spatial contextual classification method is proposed based on the high-order conditional random field model.With the consideration of improved edge structured based local dependence and distributed space contextual information,a smooth classification result can be achieved form individual classification result.Experiments and comparisons on two public standard datasets show that the proposed method can obtain more complete and higher-accuracy classification results.In the comparison of key indicators average f1-score,which is the average quality value of two datasets,the proposed method reached 85.49% and 85.57% respectively,outperforming all the other comparative methods.(3)Research on urban scene point cloud based individual objects extraction method.The result of point cloud classification only contains the class category information in the scene,lacks the corresponding individual objects information,which need further processing for practical application.In this paper,the super-voxel segmentation method is used to segment the point cloud data.Then,we analyze the impact of the scale selection on the segmentation results.The individual objects are derived by clustering the classified super voxel entities with link-chain method.The experimental results show the effectiveness of the link-chain clustering method.