Densenet 3D for Violent Action Recognition in Surveillance Video Sequences
Jorge Luis Suaña Chambi, Juan Carlos Cáceres, César Beltrán · 2022
Automatic fight detection in video sequences is an important topic for surveillance systems. The use of machine learning techniques made possible the better detection of fight,however, the models have difficulties in identifying fights in asequence of events in real time, due to the multiple degrees of freedom in the video capture such as: lighting, focus, resolution etc. Therefore, in this work, we propose a model based on the3D Densenet Convolutional Network with space-time learning features for the detection of violent actions in surveillance videos sequences. We validatedour model with four datasets, three commonly used datasets aimed at fight detection and a newdata set collected from surveillance videos. Experimentation has demonstrated that our deep learning approach can discriminate fight scenes with significantly high accuracy and it is superior than other previous studies.