Physical aggression presents a pervasive challenge worldwide, disrupting various aspects of individuals’ lives and societal functioning. This phenomenon stems from difficulties in emotional regulation, interpersonal conflicts, and socio-economic factors. Women and minors are particularly vulnerable demographics, facing high rates of violence in both intimate and public settings. Despite concerted efforts, violence detection remains a crucial issue with implications for public safety and social well-being. In this study, the challenge of violence detection in videos is addressed utilizing a combination of pre-trained VGG19 and Bi-LSTM layers. While promising results have been demonstrated in previous research utilizing VGG16, the potential effectiveness of VGG19 in this context has not been thoroughly investigated, although a larger number of convolutional layers should mean a better understanding of the scene. Moreover, the use of Bi-LSTM layers has been shown to be superior to the use of LSTM layers by up to 3%. A broad range of hyperparameter combinations is explored to optimize model performance. Positive results are obtained through experiments, with accuracies of 97%, 90%, and 73% achieved on the Hockey Fights dataset, Violent Flow Dataset, and RWF-200, respectively. Although the model does not surpass state-of-the-art approaches utilizing VGG16, it exhibits promise when compared to other proposals within the field. Overall, our study contributes to advancing violence detection methodologies, emphasizing the importance of leveraging deep learning techniques for improving public safety and social well-being.