SCHOOL VIOLENCE BEHAVIOR RECOGNITION USING SKELETON-BASED DEEP LEARNING METHODS.

Abstract

School violence is a critical social issue that demands effective automated surveillance systems to ensure student safety. Traditional video-based methods often struggle with high false-alarm rates in complex environments such as crowded classrooms. This thesis proposes a framework for school violence behavior recognition using skeleton-based deep learning methods. We leverage YOLOv8-Pose for human pose estimation and extract spatio-temporal features using a GCN combined with a BiLSTM model. To enhance model generalization, we constructed a hybrid dataset comprising over 4,000 videos processed into skeletal coordinate sequences, yielding more than 20,000 samples, by merging four datasets. A dynamic post-processing system integrating interaction and velocity filters is introduced to minimize false positives caused by daily activities. Experimental results demonstrate that the proposed model achieves an overall accuracy of 80% for violence detection, proving its feasibility for real-time deployment in educational environments.