Abstract
In recent years, surveillance cameras are widely deployed in public places,and the general crime rate has been reduced significantly due to theseubiquitous devices. Usually, these cameras provide cues and evidence aftercrimes conducted, while they are rarely used to prevent or stop criminalactivities in time. It is both time and labor consuming to manually monitor alarge amount of video data from surveillance cameras. Therefore, automaticallyrecognizing violent behaviors from video signals becomes essential. In thispaper, we summarize several existing video datasets for violence detection andpropose a new video dataset with more than 2,000 videos captured bysurveillance cameras in real-world scenes. Also, we present a new method thatutilizes both the merits of 3D-CNNs and optical flow, namely Flow GatedNetwork. The proposed approach obtains an accuracy of 86.75% on the test set ofour proposed RWF-2000 database.