Violence Detection Based on Multisource Deep CNN with Handcraft Features
Narenthirakumar Appavu, C. Nelson Kennedy Babu
Abstract
Today's latest video surveillance technology has recently been used to monitor human interactions in automated processing systems. They play an important part in security matters. There are many contests in unique between violent and non-violent. Supernatural activities Like crowded atmospheres and camera view. In this article, we proposition a deep novel violence against this structural system Built on specific features derivative from craft methods and through which violence is detected. These features are related to the representation of the image, the appearance of the image, and their motion speed, and are fed as input to a neural network (CNN), which transforms them into spatial, temporal feature, and feature streams, trained a network through this spatial stream to recognize contextual patterns in each frame of video. This temporal stream consisted of three consecutive frames for learning each dynamic pattern of violent behavior. Differential measurement of optical flow. Furthermore, we familiarized a discriminatory feature with a new kinetic energy image to signify violent acts distinct from others in a spatio-temporal stream. The approach incorporates dissimilar features of violent actions by combining the significances of these streams. CNN is also called violence and trained label. They include hockey, movie and VP datasets that are crowded and not. These tentative results determine that the proposed in rapports of accurateness and dispensation time, the violence detection technique outperformed the prior study's results.