人群运动--Scene-Independent Group Profiling in Crowd

Scene-Independent Group Profiling in Crowd CVPR2014
http://www.ee.cuhk.edu.hk/~jshao/CUHKcrowd.html
https://github.com/amiltonwong/crowd_group_profile

crowd 由 groups 组成，这里我们对 groups 属性进行分析，提出几个可以定量分析的描述算子。
Groups are the primary entities that make up a crowd.

同类之间我们使用 collectiveness, stability, and uniformity 来描述类成员的行为
Intra-group properties, e.g. collectiveness, stability, and uniformity, denote internal coordination among members in the same group.

不同类之间我们使用 conflict 来描述类之间成员的行为
Whilst inter-group properties, e.g. conflict, reflect the external interaction between members in different groups

我们的目标是 characterize and quantify crowd 中的 groups ，对其行为进行分析，人群场景的理解。 一个 group 不仅仅是个体空间上的聚合，同时也是一个动态的单元，展示出一些类内 和类间的属性，可以用这些属性来比较不同的 crowd system

3 Profiling Group Properties
这里我们将一个 group 看作一组个体具有共同的目标和 行为的一致性， with a common goal and collective behaviors
给定一段视频，我们可以将视频中所有的 group 检测出来。 每个group 包含 一个 由 KLT特征点跟踪器检测得到的 跟踪轨迹 tracklets 集合，对于每个 检测到的 group，我们希望可以提出了一些 可视化的描述子用于表示 这个 group的相关属性

3.1. Collective Transition Prior
人群中 group 的精确检测是很有挑战性的工作。这里我们假定场景中人的运动内在属性，我们用一组有限的 Collective Transition (CT) priors 来表征
这些 priors 可以在 检测 group 的同时被发现。有了这些 priors 的约束，我们的group 检测就更 robust，同时我们也可以从这些 priors 里推导出一些 group 的属性

对于每一个 pedestrian group 都有一个特定的 CT prior，可以从视频段中被发现。对于n个跟踪轨迹 tracklets ，我们假定存在 m个 Markov chains，其中 m < n and m is inferred automatically。 每个 Markov chain 有如下形式：

其中 where the continuous observation z evolves by a transition matrix A 。 Gaussian noise v_t ∼ N (0,Q) is assumed between transition
We denote Θ = {A,Q,µ,Σ} as the parameters of the chain.
A represents the CT prior

下面我们来说说怎么在检测 group的同时发现这些 prior

3.2. Group Detection by Collective Transition
问题的关键是在 视频段中 找到符可以很好拟合 the discovered priors 的 pedestrian groupings
The missing data of z k can be inferred with EM

3.3. Group Descriptors for Crowd Scenes

Collectiveness: 描述个体运动和群体运动的吻合度
The collectiveness property indicates the degree of individuals acting as a union in collective motion.

Stability: group 随着时间的变化 其内在拓扑结构是否保持不变，
The stability property characterizes whether a group can keep internal topological structure over time.
stable members tend to (1) maintain a similar set of nearest neighbors; (2) keep a consistent topological distance with its neighbors through out a clip; and (3)amember is less likely to leave its current nearest neighbor set.
定义了三个 stability descriptors

Uniformity: 个体分布的均匀性
Uniformity is an important property for characterizing homogeneity of a group in terms of spatial distribution.
这里使用图论里的 graph cut , We quantify uniformity by inferring the optimal number (c ∗ ) of graph cuts on the K-NN graph.
A higher c ∗ suggest a higher degree of non-uniformity.

Conflict: group 之间的冲突性
The conflict property characterizes interaction/friction between groups when they approach each other.

4 Applications and Experimental Results