Comparison of Transfer Learning Behaviour in Violence Detection with Different Public Datasets

The detection and recognition of violence have been area of interest to research, mainly in surveillance, Human-Computer Interaction and information retrieval for video based on content. The primary purpose of detecting and recognizing violence is to automatically and in real-time recognize violence. Hence, it is a crucial area and object of several studies, as it will enable systems to have the necessary means to contain violence automatically. In this sense, pre-trained models are used to solve general problems of recognition of violent activity. These models were pre-trained with datasets from: hockey fight; movies; violence in real surveillance; and fighting in real situations. From this pre-training models, general patterns are extracted that are very important to detect violent behaviour in videos. Our approach uses a state-of-the-art pre-trained violence detection model in general activity recognition tasks and then tweaks it for violence detection inside a car. For this, we created our dataset with videos inside the car to apply in this study.