基于轻量化时间注意力增强的视频异常检测

Weakly supervised video anomaly detection based on lightweight temporal attention enhancement

基于弱监督学习的视频异常行为检测算法使用较少的人工注释就能获得较大的性能提升,由于视频分段的影响,时序特征增强对于视频异常检测任务非常关键。现有的方法不能同时兼顾时间注意力增强与模型计算效率,因此,该文引入了一种基于轻量化时间注意力增强的视频异常检测算法,并在此基础上生成与异常检测任务相关的特征。此外,通过所设计的排序函数优化训练异常分数,能使每个片段的异常分数更加准确。相较于传统的Real-World算法,该方法在ShanghaiTech和UCSD Ped2数据集上的准确率分别提升了12.46%和13.03%,验证了其有效性。

The video anomaly detection algorithm based on weakly supervised learning achieves significant performance improvement with minimal manual annotation.Due to the influence of video seg-mentation,enhancing temporal features is crucial for video anomaly detection tasks.However,existing methods cannot simultaneously consider both temporal attention enhancement and model computational efficiency.Therefore,this paper introduces a video anomaly detection algorithm based on lightweight temporal attention enhancement,which generates features related to anomaly detection tasks.Furthermore,utilizing the designed ranking function to optimize the anomaly scores can further improve the accuracy of anomaly scores for each snippet.Compared to traditional Real-World algorithms,the proposed method achieved AUC improvements of 12.46%and 13.03%on the ShanghaiTech and UCSD Ped2 datasets,respectively,validating its effectiveness.