Violence Detection Using EfficientNet-B0 and SA-Separable Convolutional LSTM
Zhenjiang Fu, Junjian Huang, Luting Zhang · 2024
Automatic detection of violent behavior is an important research task in the construction of modern smart cities. To address this huge challenge, we proposed a deep learning model that combines the lightweight network EfficientNet-B0 and the deep separable convolutional long short-term memory (SepConvolutional LSTM) network. The model adopted a background subtraction algorithm to reduce the model's computational cost. Finally, the model integrated the self-attention mechanism into the deep separable convolutional LSTM module, thereby enhancing the model's ability to capture key information in videos. The model was applied on the largest violent dataset and its effectiveness was evaluated. The final results show that our model has superior performance in terms of computational complexity and accuracy.