Lutfiah Tri Syahyaningsih, Salsa Dillah, B. Muhammad Fajar, Jumadi Mabe Parenreng, Dyah Darma Andayani, Fhatiah Adiba, Andi Baso Kaswar
Bullying demands efficient automated detection, but existing computer vision methods often struggle with subtle gestures and complex temporal dynamics. This research proposes a hybrid fusion architecture integrating kinematicgeometric features from MediaPipe Holistic (pose, hands, and engineered angles) with spatial visual features from MobileNetV2. The resulting 1511-dimensional vectors are processed using a Bi-LSTM network reinforced by a SelfAttention mechanism. Evaluated on 1, 000 video samples across four classes (hitting, pushing, kicking, and non-violent) using a zero-leakage protocol, the model achieved 89% accuracy. An ablation study confirmed this multimodal fusion significantly outperforms kinematic-only baselines (73.50 %). Furthermore, operational testing demonstrated real-time viability with an inference latency of 1 3 0. 4 3 milliseconds per 30-frame sequence (∼ 230 FPS). However, cross-domain evaluation yielded 34.50% accuracy, indicating susceptibility to background bias (domain shift) and close-contact occlusion. the proposed system serves as an effective, real-time early warning tool for specific environments, paving the way for future research in domainagnostic action recognition. © 2026 IEEE.
State University of Makassar, Departement of Informatics and Computer Engineering, Makassar, Indonesia
