Real-Time Violence Detection Across Domains: Evaluating Model Robustness in Varied Surveillance Contexts

As urban environments become increasingly complex, ensuring public safety through effective surveillance has become a critical challenge. Traditional monitoring systems, which heavily rely on human operators, are prone to fatigue, delayed response times, and subjective interpretation. Automated violence detection powered by machine learning (ML) and deep learning (DL) algorithms presents a transformative solution by enabling real-time identification of violent behaviors.This paper provides a comprehensive analysis of recent advancements in violence detection, focusing on deep learning architectures such as Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), hybrid CNN-RNN models, and Vision Transformers (ViTs). We highlight the strengths and limitations of these approaches, particularly their real-time applicability and deployment on resource-constrained devices.A key contribution of this study is an in-depth examination of generalization challenges in violence detection models. We empirically evaluate a state-of-the-art CNN-LSTM model trained on a sports-based dataset (Hockey Fight dataset) and test its performance on a real-world surveillance dataset (AIRTLAB dataset). Our findings reveal a significant ac curacy drop from 89.5% to 56%, emphasizing the limitations of dataset-specific training. Through detailed error analysis, we identify key factors affecting model generalization and propose strategies such as multi-dataset training, domain adaptation, synthetic data augmentation, and hybrid model architectures to enhance robustness in diverse real-world scenarios.By addressing these challenges, this study provides valuable insights into developing more adaptable and reliable real-time violence detection systems, paving the way for improved public safety measures in high-risk environments such as schools, stadiums, and city centers.