Early event detection based on dynamic images of surveillance videos

摘要

Early event detection is intended to flag an event as early as possible, but before it terminates. It is critical for detecting on-going events in many applications such as spotting dangerous or criminal incidents. In this letter, we address this issue by converting video clips of a proceeding event into so-called dynamic images, which are capable of simultaneously capturing both the appearance and temporal evolution of the occurrence. By using the dynamic images of two categories of video clips (complete target event as the positive set and random segments that do not contain the target event as the negative set), we propose a novel method for training a detector based on deep learning techniques. The approach is capable of scoring partial events by monitoring the degree of event completion as it monotonically increases toward termination. In particular, we discuss experiments on the detection of humans falling and the breakout of a fighting. Experiments on several datasets illustrate the effectiveness of the proposed method.