Computing Solution for the Recognition of Basic Actions of Violence in Real Time, from the use of Convolutional Neural Networks, Video Sequences and High Performance Computing

Among the most important social problems that Peru is currently facing, the high rates of domestic violence (violence against women and children) and criminal acts (assaults and robberies) frequently carried out on the streets stand out. In this work, a real-time tool for the detection of two types of violent actions is proposed: kick and punch. This research proposes to use the CNN called YOLO (You Only Look Once). The methodology involves Supervised Learning and Transfer Learning techniques since there is a small batch of data for training. In addition, a database of 1000 images (Frames) has been generated from 90 video sequences showing violence, which were obtained from the internet (YouTube) and by own elaboration (video recording). Taking into account that conventional computers have many limitations and that this type of work requires a large computational capacity, the processing was carried out in the IIAP "Manati" Supercomputer, in this way the tool can be run in real time. This computer solution achieved 84% accuracy, to detect two main acts of violence: punch and kick; which shows an appropriate result for the use and application of the tool. The results are promising and show that the proposed strategy is adequate to reach a solution.