Understanding the behavior of humans has never been an easy task for computing systems. Several deep learning and machine learning methods have been employed to understand this, with varying degrees of success. In order to investigate subjects like action recognition and the identification of human emotions in photos, most human activity analysis research employs images or videos. For this research, information is obtained from news portals, social networking sites, and other sources. In this study, we have searched, collected, and used videos from several social media platforms, such as Instagram reels, YouTube shorts, and Twitter postings, which are categorized as either criminal or safe content. Our goal was to identify the best video categorization model architecture for use in future research and product development. With an accuracy rate of 94%, the Single Frames Video Classification Model outperformed other models in terms of performance.