Key points are not available for this paper at this time.
Acts of violence may occur at any moment, even in densely populated areas, making it important to monitor human activities to ensure public safety. Although surveillance cameras are capable of detecting the activity of people, around-the-clock monitoring still requires human support. As such, an automated framework capable of detecting violence, issuing early alerts, and facilitating quick reactions is required. However, automation of the entire process is challenging due to issues such as low video resolution and blind spots. This study focuses on detecting acts of violence using three video data sets (movies, hockey game and crowd) by applying and comparing advanced ResNet architectures (ResNet50V2, ResNet101V2, ResNet152V2) with the use of the bidirectional gated recurrent unit (BiGRU) algorithm. Spatial features of each video frame sequence are extracted using these pre-trained deep transfer learning models and classified by means of an optimized BiGRU model. The experimental results were then compared with those achieved by wavelet feature extraction approaches and other classification models, including CNN and LSTM. Such an analysis indicates that the combination of ResNet152V2 and BiGRU offers decent performance in terms of higher accuracy, recall, precision, and F1 score across the different datasets. Furthermore, the results indicate that deeper ResNet models significantly improve overall performance of the model in terms of violence detection scores, relative to shallower ResNet models. ResNet152V2 was found to be the ultimate model across the datasets when it comes to a high degree of accuracy in detecting acts of violence.
Merit et al. (Mon,) studied this question.