GITAAR-GIT based Abnormal Activity Recognition on UCF Crime Dataset

Data and computer science advancements in recent years have greatly benefited people’s day-to-day lives. On the other hand, criminals are embracing new technologies to bolster and broaden their operations. There is a great deal of promise in applying the Deep Learning (DL) paradigm to the analysis of highly structured data. However, the availability of public datasets in the crime detection area is low and task-specific, making it difficult to study and develop DL-assisted solutions that are both accurate and robust. This study aims to adapt the popular UCF-crime dataset for use with video subtitling and propose a hybrid model GITAAR (Generative Image-totext Transformer for abnormal activity recognition), a new transformer-architecture for video-caption-generation. In this paper, UCF crime dataset that compares a recently suggested video captioning system against a large number of state-of-theart methods, describing both the qualitative and quantitative aspects.