Crowd Abnormal Detection Using Two-Stream Fully Convolutional Neural Networks

The safety of public places and the monitoring of abnormal events have always been important, and there are a lot of crowd abnormal events detection methods have been proposed in recent years. Most of these proposed methods design complex hand-crafted features to represent the surveillance videos, although these methods can achieve satisfactory results, however the disadvantages of the immense computational complexity and affected by scene variations are obvious. This paper exploits Fully Convolutional Neural Networks (FCN), which has been proved to be powerful in image processing, to extract the features of videos. In order to get more useful appearance information and motion information, both the individual video frame and the optical flow of a pair of consecutive video frames are used as the input to a pre-trained FCN. In this paper, the two-stream CNN for video classification is changed to be two-stream FCN, then we utilize a novel method to compute abnormal coefficient based on the feature map from FCN. Our method is validated on abnormal detection benchmarks, and the results show it is competitive with the state-of-the-art methods.