Multi-dimensional LiDAR Point Cloud Segmentation Method Research Based On Manifold Theory

Airborne Light Detection and Ranging (LiDAR) is a fast evolving technique for aerial mapping and surveying, and also an important technical method for obtaining high-precision terrain surface. The point cloud data has become a standard necessity for surveying activities increasingly. Over the past decade, with advances in the technology of hardware-related sensing equipment, the present point cloud data are evolving towards high density, high spatial precision, and high attribute dimensionality. Compared with sparse data constraintly suitable for large-scale terrain generation before, technological advances enable laser scanning the ability to be more precisely descriptive for ground feature observation. Its successful application has been widely embedded in digital city modeling, vegetation classification, traffic information extraction, and environmental change studies.LiDAR data segmentation is a key element in automated point cloud data extraction and reconstruction process, and it has always been the concern and the focus of ongoing research by experts from photogrammetry, remote sensing, machine learning, and computer vision. Therefore, the study of efficient laser point cloud data segmentation method for automated data processing is of important significance.In this paper, aiming to overcome the shortcomings in the current laser data segmentation methods which require prior knowledge and imported geometric constraint model, and human intervention and experimental parameter adjustment, the research utilizes rich non-spatial attribute information, mines the different embedded physical properties contained in the scanned object by clustering multi-dimensional spatial data. By using mathematical manifold concepts the multi-dimensional feature space is processed by nonlinear clustering procedure for data points converging, pattern clustering and mode partitioning. The proposed point cloud data segmentation method also includes adaptive adjustment for bandwidth parameter and improved locality sensitive hashing structure for nearest neighbor search to optimize completion of the data processing. The main contents of this paper are as follows:1. The principle of laser scanning technology and the main components of current airborne laser radar system are introduced. Then analysis of the present point cloud data related segmentation research results and methods is reviewed, and the implementation and effect of each method was described. At last, the main research direction of this article is pointed out.2. The basic concepts used from manifold theory in this paper are introduced, and a procedural analysis of the mean clustering algorithm in vector space is studied. Extending the original algorithm to the nonlinear space, an analysis of the algorithm's properties and performance convergence is researched, also specific Riemannian manifolds for the next step is described.3. An analysis of the fixed bandwidth parameter setting problem contained in original mean clustering process is delivered, and a distance function between data points in multidimensional space which is defined by the potential suppression function is introduced. By shrinking process in the potential function, the bandwidth parameter update process can ensure that the improved approach with respect to the original method has faster convergence speedup. Experimental results indicate that the proposed improvement could divide laser point cloud data adaptively better.4. The issue of nearest neighbor search in the clustering process is studied, and a Gaussian extension method based on the locality sensitive hashing is proposed. The method first carry out a Gaussian cube partitioning, and then establish a multi-layer locality sensitive hashing virtual data indexing structure. Compared with the conventional sequential search, a substantial increase in the efficiency of the querying is unearthed.5. Several comparative experiments using the real scene benchmark data provided by ISPRS and NASA are carried out under the point cloud segmentation processing framework. The segmented results are evaluated with a reference of field data for accuracy and statistical analysis. The experiments demonstrate that the new proposed manifold theory-based point cloud segmentation processing framework can deal with obvious man-made objects such as buildings, and achieve a good discrimination for vegetation as well, while with high efficiency. The outcome proves the proposed method makes full use of multi-dimensional non-spatial information in point cloud data segmentation processing.The main contributions in the paper are reflected as:1. A manifold based nonlinear clustering algorithm is proposed on space metrics to solve the non-parametric multi-dimensional data segmentation issue. And the approach achieves a joint process unifying spatial and non-spatial properties for clustering laser scanning point cloud.2. A potential suppression function based method is proposed to define relationship between data points. And it overcomes the fixed bandwidth problem for the original clustering process which needs human intervention and experimental selection. The proposed method improves an adaptive clustering parameter selection for point cloud data, which reduces the processing time to the only 1/3 of the original fixed bandwidth method.3. A Gaussian locality sensitive hashing algorithm is proposed to set up a Gaussian partitioning for point cloud data preprocessing. In searching nearest neighbor for point cloud segmentation, the retrieval efficiency is significantly improved more than 10 times as opposed to the traditional linear neighborhood approach when the error rate is relatively low.