Video Violence Detection Using LSTM and Transformer Networks Through Grid Search-Based Hyperparameters Optimization

Abstract

The security system in public places can be improved by automatically detecting violence. Deep learning has recently gained popularity as a solution to classification problems, which improves the effectiveness of violent video detection. The authors extracted the features using a pretrained network, such as InceptionV3. To maximize the performance for violent video detection, the Grid Search approach was adopted to search for the optimal hyperparameter. The main goal is to evaluate how well LSTM and Transformer networks classify videos. The results show competitive performances in identifying violent videos, with the state-of-the-art methods. On the Hockey, Crowd, and AIRTLab datasets, LSTM outperformed Transformer with AUC scores of up to 0.976, 0.934, and 0.86, respectively.

Physical Violence Detection System to Prevent Student Mental Health Disorders Based on Deep Learning

“…However, it has shortcomings when analyzing videos' temporal (object movement) (Patel, 2021). So, LSTM is proven stable and robust for modeling long-term dependence on temporal change processes (Soeleman et al, 2022). In contrast to LSTM, convLSTM can encode spatial and temporal changes by utilizing convolutional gates inherent in its architecture.…”