Planar Region Segmentation Of Coloured Point Cloud Data Based On Architecture

Segmentation is a method which groups points based on certain similarity. This is required for information extraction from unstructured laser point cloud data. Many studies have been done on segmentation of point cloud data. The algorithms which are designed to extract planar surfaces, most commonly found surface in man made objects, group points exploiting the mathematical representation of the planar surface. This is because point clouds do not have any explicit information about the object except its 3D positional information. Recently, laser scanning systems also provide colour information as Red, Green and Blue (RGB) channels to each point in addition to the 3D coordinate in simultaneous capturing process. Likewise, the improvement of modern 3D scanning and modeling techniques also promoted a new development of segmentation of coloured point cloud data. On the other hand, the existing algorithms may big amount of calculation and is sensitive to the noise ,so this research is motivated to develop a segmentation algorithm is less sensitive to the noise which is able to create planar surface and improve the algorithm performance utilizing both location and colour information.The optimizational segmentation methods proposed in literature are reviewed following the study of colour information and Random Hoffman Transform. A segmentation strategy is first devised in such a way that geometrical and colour information are combined in a single step process. Then we used Random Hoffman Transform for improving performance of algorithm and the quick detection speed in seed planar creation process. The proposed segmentation algorithm consists following steps. Firstly, k-d tree data structure is prepared to support neighbourhood finding operation; secondly, RGB colour is transformed to chosen colour space (CIEL*a*b*); thirdly, seed planar is detected through Random Hoffman Transform ;finally, region growing approach based on geometrical similarity and colorimetrical similarity starts. This basically has seed plane selection and plane growing steps. A vector median filtering process is employed as pre-processing step to remove small details and noise in colour.In order to test the performance of the developed algorithm,we selected terrestrial laser scanning datasets from an old farm house of Germany and processed with optimised parameters. The reference segments are prepared interactively considering various object surfaces present in the data. A performance evaluation framework is defined to compare the results with the reference segments. This includes various metrics indicating the number of instances of correctly detected segment, over-segmentation, under-segmentation and noisy segments including the geometrical accuracy of the extracted planes. The performance is also evaluated by reducing the point density of the datasets.The results are first evaluated based on visual examination. Then, more detailed quantitative evaluation is performed. The combination of geometrical and colour information has been able to produce more meaningful segments and zoning more reasonable and more precise. However,the use of colour information has some adverse effects on the segmentation. The variation of colour on object surface, effect of shadow and the presence of additional objects tend to create inappropriate segments.In a word, this paper based on the coloured point cloud data in the terrestrial achieves better the recognition of building types by using CIEL*A*B space ,Randmize Hough transform and region growing approach based on geometrical similarity and colorimetrical similarity .