Crime Scene Surveillance with Deep Learning

The ability to automatically detect violence in videos is a critical task for various applications, including surveillance, content moderation, and public safety. In this paper, we present a comprehensive dataset for automatic violence detection in videos. The dataset consists of 350 video clips, categorized as either “non-violent” or “violent,” and is designed for training and testing violence detection algorithms. To address the challenge of false positives caused by fast movements and similarities between non-violent and violent behaviors, the non-violent clips are specifically recorded to include actions such as hugs, claps, exulting, and gesticulating. The dataset is organized into two main directories, “non-violent” and “violent,” with subdirectories “cam1” and “cam2” containing clips recorded from different perspectives using two different cameras.