Night vision improvement is gaining increased importance in the domain of video-based driver assistance systems. In comparison to conventional low-beam headlights, the active infrared night vision system of Bosch provides an extended visibility at night without blinding oncoming traffic. Since night sceneries in the automotive environment are characterized by uncontrolled and strongly varying light intensities, high-dynamic-range (HDR) video cameras with an image sensor in modern, nonlinear CMOS technology are applied in order to acquire the outperforming wide luminance range of at least five decades of brightness difference between darkest and brightest object. The downside of using these HDR cameras is that their performance is especially critical if exposed to glaring situations, which, for example, might be caused by uncontrollable illumination in typical street sceneries due to oncoming traffic or reflections at traffic signs. In these situations, unwanted and disturbing stray light artefacts might occur in the acquired images which mainly arise from reflections and scattering due to mechanical defects, impurities or other effects of the image sensor or lens. Therefore, the night vision cameras are currently tested for stray light by means of subjective visual inspection according to a failure catalogue which represents one of the most important steps in the quality control among other optical inspection methods. To guarantee a realistic test in a glaring situation, the camera is shifted manually in dark surroundings with respect to an external light source which is mapped as a bright spot on dark background at different positions in the acquired images. In order to improve repeatability and reproducibility of test results, to enhance test coverage and to reduce inspection error rates at acceptable inspection times, the main focus of the thesis lies in the automation and development of an objective stray light inspection procedure.

After introducing an overall inspection framework, the first part of the thesis describes the development of a motorized concept which replaces the current manual swivel unit in order to realize automated spot positioning and image acquisition. For the sake of determining the spot position in different regions inside and outside the image field, approaches such as the weighted centroid algorithm, affine transformation in combination with a Gauss-Helmert adjustment as well as extrapolation are implemented. In a second step, an algorithm based on supervised machine learning is presented which aims to objectively detect different types of stray light artefacts in the acquired images. This algorithm is composed of a combination of specific approaches for image preprocessing, feature extraction and classification using a random forest ensemble classifier. The task of the proposed procedure is to separate the inspected images into two distinct classes: those that indicate any type of stray light artefact (niO, defect) and those without any artefacts (iO, non-defect). While the steps of image preprocessing and feature extraction are explained by the concrete example of wide beam artefacts, the performance of the classification is evaluated with respect to different types of stray light artefacts by means of a k- fold cross-validation procedure. In the scope of performance assessment, the out-of-bag estimation error as an internal error measure of random forests is compared with the receiver operating characteristic based on predicted artefact probabilities. Finally, the thesis discusses two different approaches for feature analysis and selection in order to evaluate individual extracted features with respect to their importance in the scope of decision making. The first procedure represents a significance test of differences between the distributions of feature values extracted from iO and niO images. In the second approach, feature selection is realized by the ranking of variable importance scores which are determined as internal estimation measures during the training phase of a random forest. The outcome of both approaches is evaluated by repeating the classification procedure based on the reduced feature subsets and by comparing the classification results with those found using the original set of features. While the significance test is rather deemed to be an approach for an a priori interpretation of each feature’s ability to separate iO from niO distributions, the outstanding advantage of feature selection based on variable importance scores is that the multivariate interactions between all features and the random forest model are taken into account.

In conclusion, the thesis demonstrates that the proposed approach for automated and objective stray light inspection enables highly reproducible and repeatable classification results at improved false negative rates and acceptable inspection times which are almost independent of subjective human perception.

Schlagwörter: objektive Serienprüftechnik; objective test engineering; Streulicht; stray light; Bildverarbeitung; digital image processing; Maschinelles Lernen; Machine Learning; Klassifikation; classification; Random Forest; Optik; optics; Fahrerassistenzsystem; driver assistance system; Night Vision; hochdynamische Kamera; high-dynamic-range camera; HDR; CMOS; Automatisierung; automation; Merkmalsextraktion; feature extraction; Merkmalsselektion; feature selection

Berichte aus der Automatisierungstechnik