基于统计学习的人群异常行为检测

Statistical Learning-based Algorithm for Anomaly Detection in Crowds

针对拥挤道路环境下的人群监控问题,提出了一种新的异常行为检测方法。该算法首先对去背景后的训练视频提取特征向量,并通过统计学习的方法建立高斯混合模型（GMM）,然后提取测试视频的特征向量,输入到GMM中计算概率密度,最后用mean shift算法判断每帧是否异常。在UCSD数据库上进行帧级别的实验,取得了16%的同等错误率（EER）,在像素级别的实验中,取得了62%的定位准确率。实验结果表明该算法能有效检测不同场景下的人群异常行为。

In order to solve the problems of crowd surveillance in complicated road environment,a novel method for anomaly detection is proposed.Feature vectors were extracted from train videos within background subtracted firstly, and a statistical learning-based GMM （Gaussian Mixture Model）model was constructed.Then feature vectors within the test videos were extracted to calculate probability values using GMM.At last,mean shift algorithm was used to determine whether a frame was normal or abnormal.On the UCSD datasets, 16% EER（Equal Error Rate） was achieved in frame-level experiments,and 62% localization rate was achieved in pixel-level experiments.The results of the simulations validated the effectiveness of the proposed algorithm.