摘要
针对现有异常检测方法难以解释异常属性的问题,本文提出基于双侧空间窗的异常检测方法。首先,在前景检测的基础上,本文对场景边界区域进行双侧空间窗采样,提取双侧空间窗特征;随后,为了提取异常事件的速度属性、相关性属性、时间差属性的提取,本文分析了双侧空间窗的时序互相关理论和实际特性,实现了异常细分属性的描述;最后为了进一步描述目标类别属性,本文使用了基于快速傅里叶变换的外观特征,利用最大间隔思想训练异常检测模型。在真实场景BEHAVE数据库的实验中,可以看出AP和AUC评价指标超出现有对比方法,而且还能在没有先验知识指导的情况下,自动识别出监控场景出入口的位置。
Confronting with the challenge of the difficult interpretation of the abnormal properties in previous works, a novel abnormal detection method based on bilateral space windows is proposed in our work. Firstly, after foreground construction, bilateral space windows are proposed to sample on the boundary of monitoring area, whose features can effectively describe the interesting regions. Secondly, in order to extract the attributions of speed, correlation and time delay, we design sequential cross-correlation measurement and analyze its theoretical and practical characteristics. Finally, we train our abnormal detection model using max margin framework, which considers both attributions of speed, correlation and time delay and additional appearance feature using fast Fourier transform. In the BEHAVE dataset with actual monitoring conditions, our method outperforms state-of-the-art methods both in AP and AUC evaluation. Moreover, even without prior, the method can automatically identify the location of the entrance and exit of the surveillance scene.
出处
《计算机科学与应用》
2019年第1期19-27,共9页
Computer Science and Application
