Research On The Technology Of Surfel-based Three-dimensional Reconstruction Of Fruitage

Aiming at the key technical problems of surfel-based fruit reconstruction, a research onthe reconstruction process of fruit is presented and a set of frame for growing fruitreconstruction is proposed by taking the field fruits as the research objects, the point-based3D reconstruction as the theoretical basis, the computation geometric method as research tooland surfel-based digital geometry processing as the research procedure. The maincontributions include:(1) Expensive computation, low efficiency and inaccuracy remain a long-standingchallenge in extraction of fruit shape from massive background points. Towards theseobjectives, an improved K-neighborhood adaptive subdivision algorithm is proposed tosimplify point clouds for a reduction of calculation; an MLS-based local surface fitting isemployed for more accurate estimation of geometric attributes and a geometricsimilarity-based augmenter (GSA)recursive process is used for rapid extraction. As a result,the optimal segmentation parameters help to reduce the number of extracted cells to3.05%ofthe original points, a23.75%down than the pure K (4%) neighborhood.(2) As for the defects of pairwise transformation in parameter simplification andredundant transformation elimination, based on geometric descriptor, homogeneoustransformation model between two coordinate systems is theoretically deduced andtransformation matrix and parameters between two neighbor scans is optimized. Theexperiment shows that oriented-splat optimal transformation needs only five parameters andone pair of correspondences to lower the computational complexity.(3) Due to the low efficiency in finding correspondences associated with feature-basedpairwise registration, complex computation and inaccuracy related to iterative pairwiseregistration, a Feature-Difference based Step-by-step Contraction(FDSC) is presented forenhancing registration efficiency by taking advantage of feature-based and iterative pairwiseregistration, and a duplicato-error metric/verification algorithm is presented for improvingaccuracy. Experiments show that with FDSC having more rapid convergence rate, thetranslation error is decreased by65％than MUMC and4PCS, rotation error by89％, and theerror value by99％.(4) Global registration inaccuracy associated with error accumulation remains a long-standing challenge in free-form fruits. Towards this objective, a hierarchical automaticglobal registration strategy is proposed to disperse accumulative error, and a cumulative errorcorrection method is employed to reduce the error transfer between the layers. As a result, theglobal translation error is decreased by55％by this method than by sequence strategy or4-parameters of fixed coordinate method. The rotation error is decreased by65％and theaccumulative error value is down by99％.(5) A point-quadtree method is employed to simplify the corresponding imagesegmentation by taking advantage of surfel splitting positions; the texton descriptor, linearZ-Order and multi-level adaptive morton coding method is employed for saving the memoryspace; the color perception, subsampling, sub-division, oversampling and the weightedaverage (PSSOW) means is employed for saving memory and improving visual effect intexton properties calculation. Results show that multi-level adaptive morton code helps toreduce the memory by1/4than general code. While preventing the texture from distortion,PSSOW makes color memory condensed by70.3％,55.5％and70.3％than direct2Dmapping, Multi-resolution quantization method and color vision model.