无重叠视域多摄像机的数据关联算法

Data association algorithm of multiple non-overlapping cameras

大范围的监控通常会用到多个没有重叠视域的摄像机,目标的出现在时间和空间上都是离散的,摄像机间的目标关联是实现大范围的目标跟踪的关键问题。提出一种基于权重二部图的关联策略,以目标为节点,结合时间约束和空间约束构造边,目标的相似度为边的权重,通过求取二部图的最大权匹配得到最优关联结果。针对刚体目标提出一种LICS特征用于计算目标间的相似度,该特征对光照和目标姿态变化都不敏感。采用真实视频和仿真方法对算法进行实验,实验结果表明有好的关联效果。

Muhiple non-overlapping cameras are required in the surveillance of wide areas, and the observations of objects are often widely separated in time and space. The key problem of tracking objects in wide areas is the data association of non-overlapping cameras. A data association algorithm based on the weighted bipartite graph was proposed. The objects were the graph nodes, the graph edges were constructed by the temporal and spatial constraints, the weights of edges was the similarity of objects, and the association result was the maximum weight matching of the bipartite graph. A feature named Logarithm Illuminance Contrast Statistic （LICS） was proposed for rigid objects to calculate the similarity of objects, which was non-sensitive to the change of illumination and posture. Experiments with real videos and simulation validate the proposed approach.