Automated Intrusion Detection for Video Surveillance Using Conditional Random Fields

In this paper, we propose a method for intrusion de-tection in a video surveillance scenario. For this pur-pose, we train a conditional random field (CRF) onfeatures extracted from a video stream. CRFs estimatea state sequence, given a feature sequence. To detectintrusions, we analyze this state sequence. CRFs areusually trained in a supervised manner. Here, we espe-cially propose a new training algorithm for CRFs basedon expectation maximization, which can be used withunlabeled data. We apply the resulting trained CRFto separate normal activities from suspicious behavior.We have successfully tested our algorithm on 169 se-quences.