Comparative Evaluation of Machine Learning Methods for Bullying Detection in Surveillance Footage

Abstract

The increasing availability of surveillance footage has led to a growing interest in automated violence and bullying detection using machine learning. This study presents a comparative evaluation of classical and deep learning models for detecting bullying behavior in real-world video data. We assess the performance of several approaches, including Support Vector Machines, Random Forest, and a CNN-LSTM architecture trained on the Real-Life Violence Situations Dataset. Among all models, the CNN-LSTM demonstrated the best performance, achieving 95% validation accuracy and an AUC of 0.98. These results confirm the effectiveness of combining spatial and temporal features for identifying aggressive human interactions in complex scenes. The findings contribute to the development of real-time, intelligent surveillance systems capable of early bullying detection.