Video Based Physical Bullying Detection with Deep Learning: Utilizing 3 Dimensional and Dynamic Vision Sensors

This study focuses on developing a video-based bullying detection model using deep learning, driven by the increasing cases of physical bullying in Indonesia. The research builds upon the Bullying10K: A Large-Scale Neuromorphic Dataset towards Privacy-Preserving Bullying Recognition, leveraging its dataset to classify six physical bullying actions: slapping, kicking, punching, pushing, strangling, and hair-grabs. This research evaluates three deep learning architectures: Convolutional 3D (C3D), Inflated 3D (I3D), and Expanding 3D (X3D). Results indicate that C3D achieves the highest overall accuracy 73.29% and f1-score 64.48% at epoch 30, demonstrating superior ability in distinguishing similar actions like strangling and hair-grabs. X3D exhibits the fastest learning speed, making it suitable for rapid model training. While I3D records the lowest accuracy but remains highly stable across training and testing phases. A key methodological enhancement involves a skip-frame system, which selects non-sequential frames to capture broader contextual motion, leading to improved feature extraction and action recognition. This research improves model performance compared to prior studies in limited resources. The findings highlight C3D as the most effective model for bullying detection, with potential for real-world implementation in surveillance and monitoring systems.