Comparative evaluation of deep neural networks for automatic detection of violence scenes within videos
Nowadays, occurrences of violent activities in public places are numerous and continuously increasing. However, most of the surveillance systems used today are not capable of autonomously recognizing and preventing these activities. The field of activity recognition, which has been extensively studied recently and is continuously evolving, can provide us with a very effective tool for the automatic recognition of violent scenes and therefore enable their prevention.
In this work, various solutions based on training neural networks are tested to document their performance and demonstrate the effectiveness of each in differentiating violent scenes from non-violent scenes. Additionally, tools are utilized to graphically represent the behavior of these solutions, to justify and confirm the validity of the obtained results.
Read the whole thesis (in italian): Thesis Giri Matteo.pdf