摘要： 提出一种结合分层隐马尔科夫模型(LHMM)与熵值的聚众事件实时检测方法。使用长宽比消除前景中其他物体的影响，以区域中的人数和总速度为观察值，分2层训练出聚众事件的LHMM。当观察值序列与模型的相似度大于设定阈值时，利用光流法计算该帧熵值，若熵值大于设定阈值，则表示发生聚众事件；否则，为非聚众事件，继续下一帧的处理。实验结果表明，该方法具有较高的识别率和较好的鲁棒性。

关键词: 长宽比, 场景分块, 分层隐马尔科夫模型, 熵, 聚众事件

Abstract: This paper proposes a Layered Hidden Markov Model(LHMM) and entropy method to detect a gathering event in real-time. It uses aspect ratio eliminates the effect of other objects in the foreground, divides the video scene into blocks of regions, and relies on number and total speed statistics as the features. The features are encoded with LHMM to allow for the detection of gathering event. When the similarity between observation and the model is greater than the setted threshold, using optical flow calculates the entropy of the frame, entropy is greater than the setted threshold, it is judged to report a gathering event. Otherwise, not a gathering event, continuing processing next frame. Experimental results show the approach have higher recognition rate and robust.

Key words: length-width ratio, scene block, Layered Hidden Markov Model(LHMM), entropy, gathering event

中图分类号: 

• TN911.73