Automated Human Violence Detection using MobileNetV2 and Bidirectional LSTM Networks

---

Abstract

A cutting-edge automated approach for identifying instances of human aggression in video footage is presented in this research study. By utilizing an advanced combination of MobileNetV2 and Bidirectional Long ShortTerm Memory (LSTM) networks, the approach guarantees a strong analysis of consecutive frames. While Bidirectional LSTM records temporal relationships and allows for a more sophisticated understanding of activities over time, MobileNetV2 is a potent feature extractor. After being carefully trained on a wide range of datasets that include both violent and non-violent examples, the model obtains excellent performance metrics. The system has remarkable skill in properly identifying violent acts, with accuracy of 94.5%, precision of 93.0%, recall of 94.5%, and an outstanding F1 score of 95.9%. With its extensive modules for training, testing, data preparation, model design, and visualization, the project offers a solid foundation for critical assessment and real-world application. Subsequent research directions will concentrate on refining the architecture, investigating group learning techniques, and enhancing real-time inference, ultimately advancing automated violence detection in many real-world contexts. There is potential for revolutionary uses of this study in public safety, monitoring, and security.

---

Keywords

Automated Violence Detection, Bidirectional LSTM Networks, Convolutional Neural Networks, MobileNetV2, Video Analysis,