Statistische Ereignisdetektion in Bildfolgen

This thesis presents a technique to detect statistically unlikely changes in noisy image sequences. Methods for outlier detection are well known in statistical data analysis. This work applies these techniques to image processing. Appropriate statistical tests are performed to identify the relevant pixels by hypothesis testing. The image sequence is represented as a separate time series for each image pixel with the assumption that at steady state the scene is static. This assumption is commonly made for many applications in surveillance and spatio-temporal measurements. The significance level related to the hypothesis test remains the only free parameter. This allows an even comparison of the algorithm's performance across different data sets. A confidence measure is calculated for each binary decision (inlier vs. outlier). Effects such as occlusion or false positives that occur for multiple outliers are controlled by an iterative extension. The algorithm was put into practice twice 1) A complete computer vision system for an industrial laser welding process control was patented. It replaces human visual inspection for mass production and improves robustness over spatially integrating sensors. 2) The algorithm has been applied to infrared image sequences in order to distinguish events caused by two separate processes. Hence heat flux parameter estimation was improved by an outlier detector module at the beginning of the estimation scheme. The technique presented has proven to be an easy-to-configure, modular, and fast tool for event detection in image sequences.