Abnormal Events Detection Using Deep Networks for Video Surveillance

In this paper, a novel method is proposed to detect abnormal events. This method is based on spatio-temporal deep networks which can represent sequential video frames. Abnormal events are rare in real world and involve small samples along with large amount of normal video data. It is difficult to apply with deep networks directly which usually require amounts of labeled samples. Our method solves this problem by pre-training the networks on videos which are irrelevant to abnormal events and refining the networks with fine tuning. Furthermore, we employ the patch strategy to improve the performance of our method in complex scenes. The proposed method is tested on real surveillance videos which only contain limited abnormal samples. Experimental results show that the proposed approach can outperform the conventional abnormal event detection algorithm which utilized hand-crafted features.