Spatio-temporal alignment and analysis of videos for industrial applications

Abstract

In this thesis, we investigate the use of signal processing and computer vision techniques aiming to provide tools for signal analysis in a context of anomaly de- tection in industrial applications. In this kind of analysis, it is common to perform some sort of comparison between different signals, which requires the use of signal alignment techniques. The first part of the thesis deals directly with the temporal alignment of signals. We consider the cases in which the signals have the same length, when is only necessary to estimate a delay that aligns the signals, or different lengths, which also requires the use of warping techniques. The signals employed in the alignment can be either a signal-of-interest, such as video or audio, or auxiliary information obtained from the system, for example the measurement of power consumption in a moving platform. We consider the use of techniques that perform camera tracking from images to obtain a temporal alignment. The second part of this thesis studies the spatial alignment of images. Techniques for the estimation of the motion field of the image pixels are applied on a standard database and on the considered industrial application, which also involves a study of the method robustness to different image characteristics. At last, the third part of the thesis deals with the identification of events in a new application. During a fluid flow process, we obtain images containing a light interference pattern. We aim to estimate positions on this pattern, in order to extract the characteristics of the fluid employed.