Search

Today, automatic image analysis is one of the basic approaches in the field of industrial applications. One of frequent tasks is pose estimation of objects which can be solved by different methods of image analysis. For comparison two of them have been selected and investigated in this project: Convolutional Neural Networks (CNNs) and a classical method of image analysis based on contour extraction. The main point of interest was to investigate the potential and limits of CNNs to fulfil the requirements of this special task regarding accuracy, reliability and time performance. The classical approach served as comparison to a state-of-the-art solution. The workpiece for these investigations was a commonly used transistor element. As database an image archive consisting of 9000 images with different illumination and perspective conditions has been generated. One part was used for training of the CNN and the creation of a so-called shape model respectively, the rest for the investigation of the extraction quality. With CNN technique two different approaches have been realised. Even if CNNs are predestined for classification this method delivered insufficient results. In a more sophisticated approach the system learns the parameters of an affine transformation including the sought-after parameters of translation and rotation. Our experiments confirm that CNNs are able to obtain at best only a medium accuracy of rotation angles (about ± 2°), in contrast to the classical approach (about ± 0.5°). Concerning the determination of translations both methods deliver comparable results, about ± 0.5 pixel from CNN and about ± 0.4 pixel from classical approach.

Optical systems for automatic visual inspections are of increasing importance in the field of automation in the industrial domain. A new application is the determination of steering wheel angles during wheel track setting of the final inspection of car manufacturing. The camera has to be positioned outside the car to avoid interruptions of the processes and therefore, oblique images of the steering wheel must be acquired. Three different approaches of computer vision are considered in this paper, i.e. a 2D shape-based matching (by means of a plane to plane rectification of the oblique images and detection of a shape model with a particular rotation), a 3D shape-based matching approach (by means of a series of different perspectives of the spatial shape of the steering wheel derived from a CAD design model) and a point-to-point matching (by means of the extraction of significant elements (e.g. multifunctional buttons) of a steering wheel and a pairwise connection of these points to straight lines). The HALCON system (HALCON, 2016) was used for all software developments and necessary adaptions. As reference a mechanical balance with an accuracy of 0.1° was used. The quality assessment was based on two different approaches, a laboratory test and a test during production process. In the laboratory a standard deviation of ±0.035° (2D shape-based matching), ±0.12° (3D approach) and ±0.029° (point-to-point matching) could be obtained. The field test of 291 measurements (27 cars with varying poses and angles of the steering wheel) results in a detection rate of 100% and ±0.48° (2D matching) and ±0.24° (point-to-point matching). Both methods also fulfil the request of real time processing (three measurements per second).

The importance of single trees and the determination of related parameters has been recognized in recent years, e.g. for forest inventories or management. For urban areas an increasing interest in the data acquisition of trees can be observed concerning aspects like urban climate, CO2 balance, and environmental protection. Urban trees differ significantly from natural systems with regard to the site conditions (e.g. technogenic soils, contaminants, lower groundwater level, regular disturbance), climate (increased temperature, reduced humidity) and species composition and arrangement (habitus and health status) and therefore allometric relations cannot be transferred from natural sites to urban areas. To overcome this problem an extended approach was developed for a fast and non-destructive extraction of branch volume, DBH (diameter at breast height) and height of single trees from point clouds of terrestrial laser scanning (TLS). For data acquisition, the trees were scanned with highest scan resolution from several (up to five) positions located around the tree. The resulting point clouds (20 to 60 million points) are analysed with an algorithm based on voxel (volume elements) structure, leading to an appropriate data reduction. In a first step, two kinds of noise reduction are carried out: the elimination of isolated voxels as well as voxels with marginal point density. To obtain correct volume estimates, the voxels inside the stem and branches (interior voxels) where voxels contain no laser points must be regarded. For this filling process, an easy and robust approach was developed based on a layer-wise (horizontal layers of the voxel structure) intersection of four orthogonal viewing directions. However, this procedure also generates several erroneous "phantom" voxels, which have to be eliminated. For this purpose the previous approach was extended by a special region growing algorithm. In a final step the volume is determined layer-wise based on the extracted branch areas Ai of this horizontal cross-section multiplied by the thickness of the voxel layer. A significant improvement of this method could be obtained by a reasonable determination of the threshold for excluding sparsely filled voxels for noise reduction which can be defined based on the function change of filled voxels. Field measurements were used to validate this method. For a quality assessment nine deciduous trees were selected for control and were scanned before felling and weighing. The results are in good accordance to the control trees within a range of only -5.1% to +14.3%. The determined DBH values show only minor deviations, while the heights of trees are systematically underestimated, mainly due to field measurements. Possible error sources including gaps in surface voxels, influence of thin twigs and others are discussed in detail and several improvements of this approach are suggested. The advantages of the algorithm are the robustness and simple structure as well as the quality of the results obtained. The drawbacks are the high effort both in data acquisition and analysis, even if a remarkable data reduction can be obtained by the voxel structure.