Handcrafted vs. Learned Features for Automatically Detecting Violence in Surveillance Footage
Arnaldo V. Barros da Silva, Luis F. Alves Pereira · 2022
For many years, methods for detecting violence in video data used features designed by humans to extract visual information from input frames for composing feature vectors and then applied machine learning techniques to assign labels to them. Recently, Deep Learning methods are highly evidenced for this task since they can automatically learn image features. Furthermore, they usually overcome the accuracy rates obtained by classical methods based on handcrafted features. This work evaluates learned and handcrafted features for classifying video frames as 'violence' or 'non-violence'. Our results showed that learned features can not always be claimed superior since some violent scenes are only detected by handcrafted features.