Human-Motion Guided Frame Selection with Adaptive Gamma Correction for Violent Video Classification

This paper proposes a human-motion guided frame selection approach for violent video classification. The human-motion features are computed by determining the frame difference within the detected human regions. Additionally, adaptive gamma correction is introduced in the motion signal to mitigate abrupt motion. The experiment is evaluated on RWF-2000 dataset by using I3D network as a classification model. The empirical results demonstrate that the proposed method outperforms the state-of-the-art methods in violent classification. Selecting more informative frames can improve the classification performance compared to traditional frame selection methods that use uniform sampling. Therefore, the proposed method enables the extraction of more informative frames.