基于双阶段协同数据增强的暴力行为视频识别算法

Violence Video Recognition Based on Two-Stage Collaborative Data Augmentation

摘要: 暴力行为视频识别是现代公共安全领域中一项至关重要的技术。合理设计的数据增强方法可以提升暴力行为视频识别精度。针对现有数据增强方法难以全面覆盖时域和空域暴力行为信息的问题，提出双阶段协同数据增强网络。在训练阶段提出时空随机裁剪策略，生成具有时空域暴力行为信息的多样化背景和动作表达，提高模型对时空域暴力行为特征学习的鲁棒性。在测试阶段通过十字区域裁剪策略，扩大裁剪视角，提高暴力行为特征区域覆盖度。在VSD2015数据集上的大量实验验证，双阶段协同数据增强网络仅视觉模态的结果超过先进方法，取得领先性能。本研究通过双阶段协同增强机制，为暴力行为视频识别任务中的数据增强方法提供新的方案。

Abstract: Violence video recognition is a crucial technology in the field of modern public security. Well-designed data augmentation methods can improve the precision of violence recognition. To address the problem that existing data augmentation methods are difficult to fully cover violence information in the temporal and spatial domains, a Two-stage Collaborative Data Augmentation Network (TCDANet) is proposed. In the training phase, a Spatiotemporal Random Crop (STRCrop) strategy is proposed to generate diverse backgrounds and action representations, which containing violence information in the spatiotemporal domains, enhancing the model’s robustness in learning spatiotemporal violence features. In the testing phase, a Cross Area Crop (CACrop) strategy is adopted to expand the cropping perspective, improving the coverage of violence feature regions. Extensive experiments are conducted on the VSD2015 dataset. The results of the two-stage collaborative data augmentation network with only visual modality outperform advanced methods, acquiring leading performance. This study provides a new solution for data augmentation methods in violence video recognition tasks through the two-stage collaborative augmentation.