Segmentation Methods For Plant Organs On Point Cloud Data

How to segment plant organs on a given so-called point cloud data set, it is one of elemental work of measuring forestry informatization. The quality of the segmentation study will directly impact on the further research. Currently, even in the research of a single plant based on point cloud data, the segmentation methods for different organs is still in its infancy, i.e., effective method has not yet be found. Moreover, so far as only a few published classification methods are concerned, they unavoidably exist a variety of deficiencies, for instance, they, at least partly, fail to effectively estimate accurate organs structure, which will result in embarrassment in forestry automatic measurement. As for the point cloud data set itself, which was generated from a laser scanner, the sample properties/features are only composed of three particles: three-dimensional coordinates, the reflected intensity and the color(RGB tricolor). Intuitively, the three particles are isolated between each other, without considering their intrinsic structure information within samples. Undoubtedly, this can not describe the concept of “shape” of the same organs in a plant.Our researches are carried out mainly from two aspects: additional feature generation and recognition of plant organs. We highlight the contributions as follow:1. Additional feature generation. By mining local information of neighbor samples, we start from the local surface structure and topology connection, and combined with manifold theory. Concretely, we proposed the concept and calculation method of local tangent plane distribution, i.e., the tangent plane and its distribution in the neighborhood to characterize the local surface structure. This feature able to effectively mine the structure information from a local space, and it will be helpful for expand the feature space of the tree on point cloud data. To validate its performance, we have tried a variety of feature generation methods to build multidimensional fused features, including local tangent plane distribution and the statistics theories of spatial scatter distribution and normal distribution.2. Supervised learning. As a powerful classifier, support vector machines(SVMs) have played important role and achieved plenty of successful application in machine learning and pattern recognition. For binary classification, they minimize experience risks and simultaneously maximize the margin between two-class points to achieve minimization of the structural risks. However, a classical SVM has to solve a quadratic programming and this means there is a time-consuming process in training stage. Considering the plane characteristics of leaves and tangent plane, we also attempt to use two plane classifiers which respectively named Proximal SVM(PSVM) and SVM via Generalized Eigenvalue(GEPSVM). Compared to standard SVM, the latter two both have a shorter training time and comparable recognition performance according to their experimental reports.3. Unsupervised clustering method. However, in part 2, the class labels of the training samples are assigned by researchers’ experience. In doing so it is not only time-consuming but also subjective assertion. Along this point, we have tried a two-stage method. In the first step, a small number of the unlabeled samples were randomly selected to do clustering analysis, forming the labels by using random projection plus Gaussian mixture model(GMM) approach. In the second step, on basis of the first step, by using cluster ensemble method to stable multiple random projection and GMM clustering results. According to the theories, random projections are sanguine about avoid the local optimization problem. It is well-known that, neither random projection nor GMM is stable to the global optimal solution. Therefore, we proposed a method called cluster ensemble, it has more stable performance in clustering compared with single random projection plus GMM method, and the labels are more realistic.4. Integrated the implementation and technology in various stages of this thesis, we present an automatic segmentation method for plant organs on point cloud data. The method starts from feature generation methods, use small random sample for cluster ensemble, use labeled sample to train classifiers, and then get all the labels of the data by a semi-supervised learning processing. By this method, we will complete the identification and segmentation task for the different plant organs on point cloud data.The data set in this thesis came from the michelia or sakura tree in the campus of Nanjing Forestry University, by Leica Scanstaion C10 laser scanner. The experiments show that the proposed automatic recognition method for plant organs on point cloud data is feasible.