USO DE DEEP LEARNING APLICADO NO RECONHECIMENTO DE AÇÕES HUMANAS A PARTIR DE VÍDEOS EM ALTA RESOLUÇÃO VISANDO IDENTIFICAR MOVIMENTOS SUSPEITOS

Abstract

The use of computer vision plays an important role for security purposes. However, the combination with deep learning techniques and convolutional neural networks are still little explored because they demand a lot of computational processing capacity. This work aims to combine these techniques in order to generate an algorithm that is capable of identifying and tracking individuals in videos, in addition to monitoring their actions with the purpose of identifying movements that could signify a criminal act, using the YOLO algorithm for identification, Kalman filter for tracking and BlazePose for movement identification. This work resulted in a 95% accuracy rate on well-defined videos and an 81% accuracy rate using video from the most popular security cameras.