基于时空信息约束的密集人群分割方法

Segmenting groups in crowd based on spatiotemporal restraint

在密集场景中,人群可以看作是由三五成群的、具有运动一致性的子群体构成的集合.为了能划分这些子群体,以便对人群的交互行为进行分析,提出了一种基于时空信息约束的分割方法.该方法首先利用背景建模和特征点跟踪,获取视频图像帧中运动目标的时空信息;然后,利用前景中群体分布的空间区域信息,将空间上相邻近的人群划分为若干个子群体;最后,通过一段时间内的运动相关性将子群体进行进一步的分割.这两种约束信息相互作用,得到具有运动一致性的子群体.通过大量实验验证测试,表明该方法可以应用于不同人群密度、视角范围等多种场景视频.

Pedestrians in crowded scenes can be seen as a collection of groups moving coherently. To segment the crowd into groups and analyze interactions between them, a method using group spatiotemporM relationships is proposed. Spatiotemporal information of crowds by modeling the background and tracking the feature points is acquired. Then the individuals are grouped according to aspatial neighborhood restraint. The groups are further divided by motion correlation over time. These two restraints work with each other effectively, and generate groups with coherent motion. Tests on many videos of real-world pedestrian scenes show that the method can be applied to a variety of scenes with different crowd densities and perspective of videos.