Registering 3D Point Clouds: An Experimental Evaluation

摘要

Four separate laser scans of a wooden box, taken from different vantage points,were examined in a laboratory setting. Visual and numerical registration methods, aimed at aligning the individual scan data with respect to a common frame, were explored. The numerical registration method aligns point clouds with triangulated elevated surfaces optimizing residual-based measures-of-fit, as outlined in this report. Essential procedures for data filtering are described including methods for shadow delineation. Data phenomena beyond common noise were observed, particularly, 'phantom points' resulting from split signals (mixed pixels). Such points interfere with determining occlusions. Results from four experiments applying numerical and visual procedures are reported.