密集人群建模中的阈值训练

Threshold training in dense crowd modeling

为了实现密集人群的跟踪采用一种建模的方法,而为了对密集人群视频数据进行建模,首先要对视频中存在的运动模式进行聚类,该方法采用了一种在线聚类的方式,而进行聚类的依据为KL阈值的大小。将视频中获取的第一个运动模式定义为初始运动模式,对于其后的运动模式,计算其与已存在运动模式间的KL阈值,判断该运动模式属于哪一类,进而完成模式的聚类。通过对于不同场景、不同条件下的视频进行聚类的分析,包括相同模式相邻位置、相同模式间隔位置、不同模式相邻位置、不同模式间隔位置以及不同模式间的混合训练,分析出KL值的特性,以及KL阈值的选择机制。

Using a modeling way to realize the tracking of dense crowded.In order to model for the video data of dense crowded,cluster the current existing motion pattern in the video firstly which uses a online method,according to the Kullback Leibler(KL) divergence.Regard the first obtained motion pattern as the initial motion pattern,for the next motion patterns,conclude the KL between them and the existed motion patterns,judge which motion pattern they belong to,then finish the clustering of the motion patterns.Through the clustering analysis in videos with different scenes and different conditions,including the same pattern in adjacent positions,the same pattern in spacing positions,different patterns in adjacent positions,different patterns in spacing positions,and the hybrid training between the different patterns,analyzing the characteristics of KL and the choosing mechanism of KL.