CNN-Based Violence Detection: Evaluation and Experiments

Given the growing reliance on surveillance cameras and the increasing need for automatic monitoring of violent activities, various deep learning-based approaches have been developed to address these challenges. In this study, we started from two existing methods for violence detection in video streams based on convolutional neural networks, aiming to evaluate their performance and to investigate potential improvements. Reproducing these methods in our experimental setting allowed us to gain deeper insights into their implementation challenges, computational efficiency, and adaptability to various scenarios. The first approach employs an enhanced 3D CNN with a compact architecture inspired by DenseNet elements for spatiotemporal feature extraction, while the second method explores a more complex and computationally intensive architecture. We assessed the strengths and limitations of these models and to identified directions for optimizing the detection pipeline for real-world applications.