Violence Recognition on Videos Using Two-stream 3D CNN with Custom Spatiotemporal Crop

Abstract

Violence may happen anywhere. One of the ways to know and overseethe violence in some places is by installing Closed-circuit Television(CCTV). The recorded video captured by CCTV can be used as proofin a law court. Violence video classification is also one of the topicsbeing discussed in deep learning. The latest violence video dataset isRWF-2000. That dataset contains violent and non-violent videos, 5 secondsduration, 30 frames per second, with the amount of 2000 videos.That publication also has the best accuracy of 87.25% by their proposedmethod. In this study, we will use a Residual Network known tohave the advantage of solving the vanishing gradient problem. Besidethat, we also implement transfer learning from Kinetics and Kinetics+ Moments in Time pre-trained data. We also test the numberof frames and the location of the sampling frame range. RGB andoptical flow inputs are separately trained with different configurations.The RGB input best accuracy is 89.25% with pre-trained Kinetics +Moments in Time, using frame location 49-149. The optical flow inputbest accuracy is 88.5% with pre-trained Kinetics, using 74 frames. Wealso try to sum the output of both inputs making accuracy of 90.5%.