基于交互信息和对应关系的摄像机网络拓扑研究

Research on camera network topology based on mutual information and correspondence

提出了一种当不同摄相机的观测结果之间的对应关系未知时,通过测量不同摄像机观测值之间的统计相关性来推断摄像机网络拓扑的新方法。具体实现是通过统计相关性的非参数化估计和未知对应关系的贝叶斯集成,将两个摄像机之间的统计相关性与每个摄像机的观测值之间的交互信息关联起来,交互信息的计算是通过计算过渡分布的熵来进行的,并采用马尔可夫链蒙特卡罗集成观测值之间的未知对应关系,同时还考虑了两个摄像机的观测值之间丢失的对应关系。基于一条模拟的和真实的道路、一个模拟的和真实的摄像机交通网络的仿真结果表明,方案能够在多模态情形下比较准确地恢复过渡分布和匹配目标。

In this paper,a new method for inferring a camera network topology is proposed by measuring statistical correlation among observations in different cameras when the correspondence is unknown. The concrete implementation is to relate the statistical correlation between the two cameras with mutual information among observations at each camera by non-parametric estimates of statistical correlation and Bayesian integration of the unknown correspondence. Mutual information is computed by calculating the entropy of the transition distribution,and Markov Chain Monte carlo is used to integrate the unknown correspondence among the observations,while considering the missing correspondences between observations in two cameras. The simulation results based on a simulated and real road and a simulated and real traffic network of cameras show that the proposed scheme can recover the transition distribution and match objects accurately in multi-modal situations.