Unsupervised Abnormal Behavior Detection for Real-Time Surveillance Using Observed History

This paper presents a novel method of utilizing observed history for detecting abnormal behaviors in surveillance applications. An unsupervised algorithm is proposed to detect abnormal behaviors and re-train itself in real-time. Motion vectors of objects are estimated using the optical flow method. Encoded feature vectors are stored in an observation matrix, abnormal behaviors can be detected by applying principal copmonent analysis (PCA) on the matrix. This method has been evaluated under both indoor and outdoor surveillance scenarios. It demonstrates promising results that this detection procedure is able to discover abnormal behaviors and adapt to changes in the behavioral patterns incrementally.