基于稀疏表示和低秩逼近的自适应异常事件检测算法

Adaptive abnormal event detection algorithm based on sparse representation and low rank approximation

针对传统异常事件检测算法没有考虑视频数据低秩特性的问题,提出了基于低秩稀疏编码模型的字典学习算法。对提取的多尺度三维时空梯度特征进行K-均值聚类。利用低秩稀疏编码模型进行每一个特征聚类的字典学习。通过迭代聚类和字典学习获取所有的正常行为模式。采用公共数据集UCSD Ped1和Avenue检测该算法的性能。与社会力(SF)、混合概率主成分分析(MPPCA)、社会力-混合概率主成分分析(SF-MPPCA)、混合动态纹理(MDT),Adam、子空间(Suspace)、稀疏组合学习框架(SCLF)7种方法对比,该文算法具有较高的正确率和较强的实时性。

A dictionary learning algorithm based on a low rank sparse coding model is proposed aiming at the problem that traditional abnormal event detection algorithm doesn’t consider the low rank characteristic of video sequences. Multi scale gradient characteristics of three-dimensional space-time are extracted and clustered using K-means clustering. Dictionary learning of every feature clustering is carried out using the low rank sparse coding model. Every normal behavior pattern is obtained using iteration clustering and dictionary learning. The performance of this algorithm is tested using two public data sets UCSD Pedl and Avenue. Compared with social force （ SF） , mixture of probabilistic principal component analyzers （ MPPCA ） , SF-MPPCA,mixture of dynamic texture （MDT） , Adam, subspace and sparse combination learning framework （ SCLF ）,the result of this algorithm is more correct and real-time.