中文摘要：

提出了一种基于上下文和稀疏编码框架的无监督异常行为识别方法。首先对图像进行稠密采样,获得稠密轨迹,并提取轨迹中心周围图像块的形状特征、R–HOG、HOF特征作为特征描述符,加强了对运动信息的描述。其次,将人体行为区域和上下文区域分割开来建立2个独立字典。再将它们组成联合字典最大化字典信息,避免了单独识别人体异常行为而忽略上下文信息所导致的漏报。最后,利用稀疏重构的方法进行异常检测,分别计算测试样本中上下文区域和行为区域的重构误差,相对重构误差为负表示为正常行为,否则判断为异常行为。在KTH行为数据集上进行对比实验,实验结果表明本文算法在不同背景下均能有效识别异常行为。

英文摘要：

A novel method for unsupervised anomaly detection based on context and sparse code is proposed invideo flow.Firstly,dense points are sampled from each frame and tracked to obtain dense trajectories.Shapedescriptors(point coordinates),appearance descriptors(R–HOG)and motion descriptors(HOF)of trajectories areproposed to represent human activities.Then,the human action video clip could be segmented into an action regionand a context region to construct a joint-dictionary,which could maximize the dictionary information and reducefalse-negative.The sparse reconstruction method is used to detect abnormal action.In this sense,two regionsreconstruction errors of test samples are calculated respectively.When the relative sparse reconstruction error ispositive,the samples would be judged as abnormal.