| 赵义正.基于标签置信度的弱监督自训练视频异常检测算法[J].高技术通讯(中文),2025,35(4):360~369 |
| 基于标签置信度的弱监督自训练视频异常检测算法 |
| A weakly supervised self-training video anomaly detection framework based on label confidence |
| |
| DOI： |
| 中文关键词: 视频异常检测； 自训练； 多示例学习； 弱监督学习 |
| 英文关键词: video anomaly detection, self-train, multiple instance learning, weakly supervised learning |
| 基金项目: |
| 作者 | 单位 | | 赵义正 | （安徽电信规划设计有限责任公司合肥 230000） |
|
| 摘要点击次数: 8 |
| 全文下载次数: 5 |
| 中文摘要: |
| 在公共安全领域，如何借助视频监控设备实现实时、高效的异常事件检测，已成为一个的重要研究课题。为此，本文提出一种基于隐式类激活特征和标签置信度的弱监督视频异常检测算法。针对正常与异常之间的界限模糊并会随着不同的场景而有所变化的问题，提出使用隐式类激活模块差异化正常和异常的类间特征表达。针对多示例学习框架引入的标签噪声问题，采用基于标签置信度感知的自训练策略，通过计算伪标签的置信度，在模型迭代过程不断提高伪标签的质量。本文算法在ShanghaiTech和UCF-Crime这2个公开数据集上的曲线下面积 (area under curve，AUC) 分别达到97.63%和86.38%。模型在制造业工厂实际场景中进行测试，实验结果表明所提算法能够有效检测视频中的异常事件。 |
| 英文摘要: |
| In the field of public safety, realizing real-time and efficient anomaly detection with the aid of video surveillance equipment become a significant research topic. A weakly supervised video anomaly detection framework based on implicit class activation features and label confidence is proposed. Given the ambiguous boundary between normal and abnormal events, which varies across different context, an implicit class activation module is introduced to differentiate the inter-class feature representations of normal and abnormal events. Furthermore, to tackle the issue of label noise introduced by the multiple instance learning framework, a label confidence-aware self training strategy is employed. This strategy computes the confidence of pseudo-labels, continuously improving the quality of pseudo-labels during the model iteration process. The proposed method achieves the AUC (area under curve) scores of 97.63% and 86.38% on the ShanghaiTech and UCF-Crime datasets, respectively. The model is tested in actual scenarios within manufacturing plants, and experimental results demonstrate its effectiveness in detecting anomalies in video data. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
