中文摘要：

提出一种可用于视频中群体异常和个体异常检测的方法，即先用快速稀疏编码算法生成字典，再用字典表示测试特征样本，并以重构误差作为目标函数进行异常判别；对于群体异常，用块匹配运动估计代替耗时的整帧光流计算，生成多尺度运动直方图，有效地减少了计算复杂度；对于个体异常，提取HNF特征，若稀疏表达的重构误差超过阈值，则用惊奇计算进行二次检测，判断其是否为噪声导致的虚警，计算字典中已包含和未包含的不同特征描述子之间的差别，若判别出虚警，则更新字典减少后续检测中噪声干扰的虚警数．实验表明，算法有效地提高了检测率，降低了计算复杂度，且易于实现．

英文摘要：

An anomaly detection method which can be used for crowd and individual anomalies in videos was proposed. Fast sparse coding algorithm was exploited to generate the dictionary that represents feature samples and reconstruction errors were used as the objective function to discriminate abnormalities. For crowd anomalies, motion estimation and block matching exploited instead of calculating time-consuming optical flow estimation in the entire frame, producing multi-scale histograms of motion and reducing computational complexity effectively. For individual anomalies, the features HNF were extracted, and if the reconstruction error based on sparse representation is beyond the predefined threshold, surprise computation is used for a second detection to judge a false alarm induced by some noise. The difference between different feature descriptors included and not ineluded in the dictionary is computed. If discriminated as a false alarm, the dictionary should be updated to reduce the number of false alarms induced by noise jamming in the subsequent detection process. Experiment results show that the proposed algorithm improves detection rate effectively, reduces computational complexity and can be implemented easily.