Tracking and Visual Quality Inspection in Harsh Environments (print-version)

This Thesis describes novel image template based methods for robust and real-time capable tracking and visual quality inspection tasks in harsh environments. The presented tracking methods include three different constitutive approaches. First, a spline-regularized template tracking method applied for weld seam tracking in industrial robotic welding environments is presented. Second, a template tracking method that relies on incremental template blending updates is presented for more dynamic scenes, allowing for template tracking in harsh outdoor environments. Finally, a novel generic tracking fusion framework that allows for combining an arbitrary number of trackers that report diverse typically not directly combinable outputs is presented. In this way, advantages of different tracking cues can be combined in order to solve highly complex tracking problems. The methodological focus of all presented tracking methods lies on robustness and successful handling of diverse disturbing effects, environmental influences, and image noise, while remaining real-time capable. Considering the task of image based quality assessment we present a visual quality inspection framework for industrial robotic welding tasks, consisting of a semisupervised method that incrementally generates panorama images of entire welding processes, and of an unsupervised classification approach that automatically detects welding defects in an on-line fashion. Both methods rely on weld seam image templates provided by an underlying tracking approach. The focus here lies on high panorama image quality including a consequent suppression of industrial image noise as well as prevention of typical blending artifacts, and on highly accurate classification results at a