Automatic detection of violence in surveillance videos is crucial for improving public safety and enabling timely intervention. As the volume of video streams increases, manual monitoring becomes increasingly unfeasible, necessitating efficient vision-based solutions. This study investigates and compares three deep learning architectures for violence detection: a lightweight Convolutional Neural Network (CNN) combined with Long Short-Term Memory (LSTM), a 3D Convolutional I3D model using RGB inputs, and a Two-Stream I3D variant integrating both RGB and optical flow. The experiments were carried out on a custom-labeled surveillance dataset, using evaluation metrics such as accuracy, precision, recall, F1 score and ROC AUC. The results show that the CNN+LSTM model offers fast inference and minimal resource usage, making it suitable for embedded systems, albeit with lower detection accuracy. The I3D (RGB) model provides better recall and a good trade-off between performance and computational cost. The Two-Stream I3D model achieves the best overall performance, particularly in terms of precision and F1-score, due to its improved motion representation. In addition, we integrate interpretability through Class Activation Maps (CAMs) to improve transparency in decision making. Although evaluations are limited to our custom dataset, the results highlight the real-time feasibility and deployability of these models in intelligent surveillance systems.