Anomaly detection algorithm complements human surveillance in that it is capable of handling both very fast and large volume of observations. In this report, it is instead used to evaluate and detect anomalous behavior in video-recorded extremely slow processes, which can be as challenging for human perception.
Morphological image processing techniques are used to separate the slowly evolving foreground from the background in a time sequence of images. By defining the corresponding metrics of the extracted foreground features and applying an Bayesian estimator to the sequence within a finite time window, the probability of the upcoming observation being an anomaly can be evaluated. Single crystal growth process with optical floating zone method is used as an example for application.