Developing a border constrained voxel-based segmentation method in 3D Lidar point cloud processing

Structures and objects located indoors and outdoors can be scanned with Lidar (Light Detecting and Range) systems and transferred to digital environments as point clouds in three-dimensional (3D) and colored. The points, which are the elements of this 3D point cloud data obtained by Lidar scanning and providing detailed information about the structures and objects, come in an irregular form without being located in an organized data structure. Nowadays, developments in Lidar technology have not only improved the quality of point cloud data (with less location error and higher resolution), but also brought enormous amounts of irregular data. The segmentation process is the process that reduces the number of data by grouping data with similar characteristics and spatial proximity, and provides more meaningful information from the data. Segmentation is of great importance for applications that require dealing with a lot of data in the field of computer vision, including 3D point cloud processing. The ability of the segmentation process to yield results on complex data in the desired features and time has been a separate challenge in the field of computer vision. In this thesis, a novel voxel-based segmentation method was developed by focusing on the segmentation of point clouds in order to make the segmentation process more successful and faster, which significantly affects the success of the application. The developed method was able to perform the segmentation process by using simple geometric features such as the inclination angles and barycenters of the planar structures formed by the local point groups on the surfaces. Within the scope of the thesis, considering the characteristics of the data sets used in the literature, similarly, three different 3D point cloud data were obtained by scanning one indoor and two outdoor environments with a terrestrial Lidar system. After the raw point data obtained were pre-processed such as reduction, clipping and noise removal according to the intended use of the data set, the segmentation reference segments of them were also prepared and three data sets were created. In addition to the data sets prepared within the scope of the thesis, two segmentation data sets were obtained from the literature and thus, a total of five data sets were used in segmentation comparison. After the data sets had been obtained, the development and improvement stages were proceeded simultaneously in two separate branches up to the stage of comparing the methods over quantitative values. One of them is the stages of coding the data voxelization technique with the eight-branch tree (octree) organization, the refitting pre-process for voxels that do not show planar feature, and the developed segmentation method. The other is the stages of determining the segmentation methods that have succeeded in the literature for comparison, obtaining or recoding them, and coding the accuracy and F1 score values calculation methods for quantitative comparison. After all these development, improvement and coding stages were completed, the accuracy and F1 score results of the segmentation outputs of the applied methods on the data sets used within the scope of the thesis were obtained and comparison analyzes were made in terms of success and working time. The method developed has 0.81 average accuracy value and 0.69 average F1 score value in an advantage in terms of segmentation success and speed compared to other methods using geometric properties of points similarly in the literature. Within the scope of the thesis, the differences in the color values of the points in the point cloud were also included in the segmentation at certain effect rates, and success was increased in indoor data with high color quality. In the scope of the thesis, the developed segmentation method was also examined as an intermediate process in raw point cloud classification with different segmentation parameter values on an indoor semantic segmentation dataset (S3DIS) consisting of a large amount of points obtained from the literature. For the classification process, the data was divided into segments by segmentation with the method developed first and feature vector was extracted from each segment. Then, classification is made using these feature vectors. The segmentation-based classification process was applied with two different classifiers as Support Vector Machine (SVM) and Random Forest (RF), seperately. The results of the classification processes were compared on the accuracy and F1 score values of the class labels of the points. According to the comparison results, the classification successes of the points in the raw point cloud were 0.76 accuracy and 0.48 F1 score for SVM, while 0.83 accuracy and 0.70 F1 score for RF. Looking at the results, the RF classifier gives better results than the SVM classifier according to the data and feature sets used.