FightNet deep learning strategy: An innovative solution to prevent school fighting violence

• Published in: Journal of intelligent & fuzzy systems 2023-10, Vol.45 (4), p.6469-6483
• Main Authors: Thao, Le Quang, Diep, Nguyen Thi Bich, Bach, Ngo Chi, Linh, Le Khanh, Giang, Nguyen Do Hoang
• Format: Article
• Language: English
• Subjects: Artificial intelligence; Datasets; Deep learning; Machine learning; School violence; Students; Violence
• ISSN: 1064-1246; 1875-8967
• DOI: 10.3233/JIFS-232480

In this study, we introduce a new method to address the pressing issue of school violence using Artificial Intelligence (AI). School violence is a critical issue that affects the safety and well-being of students, teachers, and the school community as a whole. Violent behaviors, such as bullying, physical assaults, and weapon use, can have long-term effects on students’ psychological health and academic performance. To reduce these issues, we developed a lightweight Deep Learning model that can be integrated into a school’s surveillance camera system to quickly detect violent fighting behaviors for timely intervention by school staff. The proposed FightNet model consists of three components: MobileNetV2 backbone, Feature Pyramid Network (FPN) neck, and Centernet Object as a Point (COaP) head. By optimizing the hyperparameters of the model to extract keypoints in image frames from the COCO dataset, we applied an LSTM model to determine the temporal dependence of actions and classify them as “fighting” or “normal” using the UBI-Fights dataset. The FightNet model achieved mAP@0.5 of 45.34% and mAP@0.95 of 55.89% in estimating keypoints, and 72.68% accuracy and 71.69% F1-score in predicting actions. Based on these results, we conclude that the proposed model can effectively address the issue of school violence.