Violence is a problem of great impact that affects the countries of the world, violence is a soft system that to date cannot be controlled, because it can manifest itself in different ways, represented by high levels of crime and delinquency. But now, there is Artificial Intelligence (AI), which uses innovative resources and will help close gaps in the Sustainable Development Goals (SDG), such as in health, traffic management, climate change and also violence. One way to apply AI is through image processing with Convolutional Neural Networks (CNN). This research evaluates the effectiveness of classifying actions of interpersonal violence such as: strangulation, grappling, kicking or punching, using CNN with Transfer Learning. First, a personalized dataset was developed using images with simulated violent actions, obtaining a total of 2000 images, which were distributed into 60% for training, 30% for validation and 10% for testing. Second, the pre-trained CNN models were trained: VGG16, MobileNetV2 and YOLOv8, subjected to 150 epochs, where the performance results were compared between them and it was determined that the best performance was for YOLOv8, obtaining an accuracy rate of 94.25% and an accuracy of 89%. In the end, a prototype web system was developed using Python with OpenCV and YOLOv8 to detect actions of interpersonal violence in real time and alert the authorities involved in a timely manner. This research will serve as a reference for future real-time video surveillance applications.