Proposing an ontology for human-related crime recognition in videos

Video surveillance has been around for a very long time and, throughout the years of its existence, it has seen a large growth in popularity. However, due to this increase in produced footage, manual monitoring is not always an option. This paper introduces a framework that groups crime and provides a classification of different relevant aspects, such as location and demeanor of the suspect. This so-called ontology of crime, in the form of a semantic tree, is then used to build a model upon the existing ResNet50 model. The proposed model achieves an accuracy of 37.7% compared to 34.4% accuracy of the regular ResNet50 model. Furthermore, an implementation where multiple frames are used to classify one instance of crime is shown. As well as an implementation of a threshold, that filters out low quality frames during the classification process