Search

Analyzing free-bending pipelines in cluttered 3-D scenarios is a vital and challenging task in industrial manufacture and maintenance. It is beneficial in understanding the structure of pipelines and monitoring pipeline spacing for accurate assembly. Most existing pipeline analysis algorithms focus on regular straight pipelines and only depend on 3-D spatial cues. In this article, we propose a novel pipeline analysis framework based on 4-D spatiotemporal pipeline representation. The framework allows multiple pipeline analysis tasks ranging from detecting free-bending pipelines in point clouds and extracting 3-D skeletons to modeling complete 3-D pipelines. The 4-D pipeline representation formulates a free-bending pipeline as a 3-D cylinder sequence and incorporates 3-D spatial geometrization of each cylinder frame into temporal continuity of the frames. The multitask pipeline analysis starts from a cylinder patch and is performed on sequential pipeline frames using spatiotemporal cues. Finally, surface registration-based optimization refines pipeline skeletons, detected point cloud, and generated complete models simultaneously. The proposed algorithm is verified on real pipeline scenarios. It can deal with challenging free-bending pipeline shapes, clutter, and occlusions. Compared with existing algorithms, our algorithm achieves more comprehensive tasks and higher detection performance. The modeling of complete pipelines reaches an overall accuracy of 0.33 mm. Finally, we present a thorough analysis concerning the applicability of our algorithm from multiple aspects. [ABSTRACT FROM AUTHOR]

3-D modeling, object detection, and pose estimation are three of the most challenging tasks in the area of 3-D computer vision. This paper presents a novel algorithm to perform these tasks simultaneously from unordered point-clouds. Given a set of input point-clouds in the presence of clutter and occlusion, an initial model is first constructed by performing pair-wise registration between any two point-clouds. The resulting model is then updated from the remaining point-clouds using a novel model growing technique. Once the final model is reconstructed, the instances of the object are detected and the poses of its instances in the scenes are estimated. This algorithm is automatic, model free, and does not rely on any prior information about the objects in the scene. The algorithm was comprehensively tested on the University of Western Australia data set. Experimental results show that our algorithm achieved accurate modeling, detection, and pose estimation performance. [ABSTRACT FROM PUBLISHER]

Autonomous robotic systems require automatic registration of data that are collected by on-board sensors. Techniques requiring user intervention are unsuitable for autonomous robotic applications, whereas iterative-based techniques do not scale well as the data set size increases and, additionally, tend toward locally minimal solutions. To avoid the latter problem, an accurate initial estimation of the transformation is required for iterative algorithms to properly perform. However, in some situations, an initial estimate of the transformation may not be readily available; hence, a method that does not require such an initial estimate nor descends into local minima is desirable. The method presented in this paper takes advantage of the multidimensional Fourier transform, which inherently decouples the estimation of the rotational parameters from the estimation of the translational parameters, to compute 3-D registration between range images without requiring an initial estimation of the transformation and avoiding problems of the classical iterative techniques. Using the magnitude of the Fourier transform, an axis of rotation is estimated by determining the line that contains the minimal energy differential between two rotated 3-D images. A coarse-to-fine approach is used to determine the angle of rotation from the minimal sum of the squared difference between the two rotated images. Due to the Hermitian symmetry introduced by the Fourier transform, two possible solutions for the angle of rotation exist. The proper solution is identified through the use of a phase-correlation technique, and the estimate of translation is simultaneously obtained. Experimental results and an extended performance evaluation illustrate the accuracy that can be achieved by the proposed registration technique on simulated and on real range images. Last, a comparison of computational stability with that of the classical iterative closest point method is presented. [ABSTRACT FROM PUBLISHER]