Detecção de ações de violência em vídeos por meio de redes neurais artificiais

Abstract

The number of actions considered violent in recent years has grown exponentially. In the most serious situations, the use of some artifact to further threaten the victim is used, making it difficult for someone to get out of the situation unharmed. Thinking about reducing this problem, public and private spaces choose to apply actions that help to promote security, among them is remote surveillance through strategically placed cameras. Bearing in mind that only cameras alone cannot be effective, methods to support remote surveillance are necessary, among them the objective of the work, which is the application of an artificial neural network (or ANN), in order to help security agents, those who operate surveillance systems, both internal and external. In general, the neural network must recognize violent contact between individuals. This detection aims to reduce the response time that a professional would have to move to the event. To achieve the proposed result, a comparison was made between two approaches - processing the video from a neural network that analyzes the video frame by frame, and also using a mix between the concepts of convolutional neural networks and longterm memory architecture. term, known as LSTM, which plays a crucial role in processing temporal information. The results obtained by the frame-by-frame neural network provided an accuracy rate of 87.61% in the detection of violent actions, while the convolutional network together with the LSTM architecture returned 90.02%, validating both approaches.