基于SFV的暴力群体行为检测算法

Violent crowd behavior detection based on stacked fisher coding

提出一种基于深度学习思想的暴力群体行为检测算法。该算法分为三个阶段,第一阶段提取一种被称为SC-ISA的时空不变性特征;在第二阶段,运用max-pooling方法与Stacked Fisher Vector Coding(SFV)结合进行特征量化;在第三阶段,运用SVM分类算法对视频进行分类检测。该算法在Crowd Violence dataset上进行了仿真实验,视频测试结果表明,其准确性和检测速度都优于对比算法。

This paper proposed a scheme totally based on deep learning idea for violent crowd behavior detection. The scheme has three phases. In the first phase,a hierarchical invariant spatio-temporal feature called stacked convolutional ISA feature is used to describe precise video motion. In the second phase,it combines max-pooling method with a two-layer stacked fisher vectors encoding strategy to make features after vector quantization more representative and discriminative. In the third phase,SVM is used in violent crowd behavior detection. It evaluated proposed method on a challenge crowd violence dataset.The experiment shows an accuracy ratio of 91. 01%. It proves that the method is better than other methods.