VioNet: An Enhanced Violence Detection Approach for Videos Using a Fusion Model of Vision Transformer with Bi-LSTM and 3D Convolutional Neural Networks

The identification of violence in real-world scenarios is imperative as it enables the detection of aggressive behavior, thereby preventing harm to individuals and communities. This is crucial for ensuring public safety, facilitating effective crime investigation, promoting child safety, safeguarding mental health, and facilitating social media moderation. Various methods, including handcrafted techniques and deep learning algorithms, can be utilized in surveillance or CCTV cameras, as well as smartphones, to enable timely detection of violent behavior and facilitate appropriate action and intervention. In this study, we introduce VioNET, a novel approach that combines a 3D Convolutional Neural Network and a Vision Transformer with Bidirectional LSTM for the purpose of accurately detecting violence in video data. Since video data is inherently sequential, the extraction of spatiotemporal features is essential to accurate detection. The use of these two deep learning methods facilitates the extraction of maximum features, which are then fused together to classify videos with the highest possible accuracy. We evaluate the effectiveness of our approach by employing three datasets: Hockey, Movies, and Violent Flow, for analysis. The proposed model achieved impressive accuracies of 97.85%, 100.00%, and 96.33% on the Hokey, Movie, and Violent Flow datasets, respectively. Based on the obtained results, it is evident that our method showcases superior performance, outperforming several existing approaches in the field and establishing itself as a robust and competitive solution for violence detection in videos.