一种基于稀疏编码模型的视频异常发现方法  

Research on Video Anomaly Discovery Based on Sparse Coding Model

作　　者：谢锦生[1] 郭立[1] 赵龙[1] 桂树[1] 

机构地区：[1]中国科学技术大学电子科学与技术系,合肥230027

出　　处：《小型微型计算机系统》2014年第4期917-921,共5页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61071173)资助

摘　　要：提出一种基于稀疏编码模型的视频异常发现方法,不仅可以检测"个体异常行为",同时也可以应用于"群体异常行为"检测.对个体异常和群体异常事件,分别提取HNF和多尺度运动直方图作为不同的特征,通过快速稀疏编码算法学习包含正常行为特征的字典,以特征关于字典稀疏表达的重构误差作为判断异常的标准,如重构误差大,则判断为异常.多尺度运动矢量直方图不仅减小了特征提取阶段的计算量,实验结果也证明了该特征在群体异常检测中的有效性,实现了帧级场景异常检测,与真实结果比较能更及时的报警异常出现的时刻.实验使用3种标准行为识别数据库,ROC曲线和AUC值证明了算法的实用性和有效性.An anomaly detection algorithm for video based on spare coding model is proposed, which can not only detect individual abnormal behavior, but also crowd abnormal behavior. The feature of HNF and multi-scale histogram of motion are extracted respec- tively as different features according to individual anomaly and crowd anomaly, and fast sparse coding algorithm is exploited to learn the dictionary including the feature information of normal behavior. Recontruction error is used as the criterion of discriminating a- nomaly. If the representation error is beyond a predefined threshold, it is discriminated as an anomaly. Multi-scale histogram of mo- tion not only reduce the computational complexity in the stage of feature extraction, also did the experiment result prove the efficiency of the feature in crowd anomaly detection, implementing frame-level detection in the scene. Compared with ground-truth result, our method can alarm the appearing moment of the anomaly more promptly. Three standard behavior recognition datasets are used in the work and the result of ROC curve and AUC value demonstrate the practical applicability and efficiency of the method.

关 键 词：异常检测 稀疏编码 时空兴趣点 视频分析 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]