Automated Intrusion Detection for Video Surveillance Using Conditional Random Fields

In this paper, we propose a method for intrusion detection in a video surveillance scenario. For this purpose, we train a conditional random field (CRF) on features extracted from a video stream. CRFs estimate a state sequence, given a feature sequence. To detect intrusions, we analyze this state sequence. CRFs are usually trained in a supervised manner. Here, we especially propose a new training algorithm for CRFs based on expectation maximization, which can be used with unlabeled data. We apply the resulting trained CRF to separate normal activities from suspicious behavior. We have successfully tested our algorithm on 169 sequences.