Enhancing Real-Time Violence Detection in Video Surveillance Using Hybrid Deep Learning Model
Mohammed Inayathulla, K Rajasekhara Rao · Journal of Wireless Mobile Networks Ubiquitous Computing and Dependable Applications · 2025
One of the crucial aspects of maintaining public safety and security in different environments is detecting violence in video surveillance (VS). Conventional systems are unable to accurately differentiate between violent and non-violent actions due to multi-factor nature and relative subtlety of violence, as well as environmental constraints. The use of advanced Deep Learning (DL) models, specifically Recurrent (NN) Neural Networks (RNN) and Convolutional NN (CNN), along with its types, such as ResNet, and bidirectional Long Short-Term Memory (Bi-LSTM) units, to address this problem. It serves as the focus of this research. To efficiently utilise both spatial (S) and temporal (T) data, the combination of ResNet50V2 architecture with bidirectional GRU and Bi-LSTM layers was employed by the suggested hybrid model. The model has a high success rate and much lower False Positives (FP) after being trained on a wide variety of real-world events. This model's computational efficiency and wide range of applications to various surveillance situations are also discussed, along with its potential for Real-Time (RT) operation. The DL architectures are an effective approach for creating VD systems that are reliable, adaptable, and scalable and it was demonstrated by the outcomes.