A Transfer-Learning-Based Deep Network for Detecting Violence in Real-Time Videos

The detection of crowd behavior is critical for public safety and overall monitoring. A smart AI-based surveillance system is therefore necessary to monitor and detect crowd behavior, which is gaining increased usage in both government and private organizations. In this chapter, we proposed a deep learning-based technique for detecting agitated and nonagitated scenarios in real-time videos. A “transfer-learning” based Inception-V3 model is employed here as deep network. To fine tune Inception-V3 for the categorization of agitated and nonagitated videos, we have added a fully connected layer with “softmax” activation function to the model having two neurons. The proposed model’s performance was assessed on three standard databases, namely Hockey Fight, Violent-Flows and Movie, concerning average accuracy measure. The proposed model achieves 97.7%, 95.43% and 98.51% classification accuracy for the aforementioned datasets, which is significantly better than other related state-of-the-art methods.