Violence Recognition with Adaptive Temporal Down-Sampling

Violence, as a critical subfield of action recognition, holds substantial practical value in real-world applications. Down sampling methods adopted by existing violence recognition approaches lack optimization for the abrupt characteristics of violent behaviors, which potentially leading to spatiotemporal feature loss. To address this issue, a novel violence recognition method incorporating adaptive temporal down-sampling is proposed in this paper. The method replaces conventional data compression with adaptive temporal down-sampling to minimize spatiotemporal feature loss during compression; additionally, to tackle the limited scale of violence-specific datasets, the proposed approach employs pre-training on a large-scale general action recognition dataset to optimize the deep learning model's convolutional layers for spatiotemporal feature extraction. Subsequently, fine-tuning is applied to retrain the fully-connected layers on the violence-specific dataset, enabling accurate prediction of violence labels. Experimental results demonstrate that the proposed method achieves a state-of-the-art recognition accuracy of 99.50% on the Hockey Fight data set, significantly outperforming the previous results, confirming the superiority and effectiveness of our approach.