一种用于多目标跟踪的特征关联方法*

在基于图像序列的多目标跟踪问题中，传统目标链算法利用目标的运动特征和面积特征对其进行关联，取得了较好的效果，但在如新目标进入视场，旧目标离开视场，多个目标靠近、融合，单目标发生分裂等复杂跟踪情况中，以上特征存在一定的失效情况。为此，提出一种计算目标形状相似度的方法，在传统目标链算法的基础上引入描述目标形状的新特征，构造了一个计算目标形状相似度的函数，可以解决形状不同的目标在靠近或交叉通过等复杂情况下的准确关联问题。实验表明，在对刚体目标的跟踪中得到了更可靠的目标关联效果。

Track multiple objects with feature-correlation algorithms

This paper proposed a method to compute objective contour feature. Conventional objective chain algorithms correlated the motion with area characteristics of the objectives to reach the goals. Nevertheless, they might not be efficient under certain circumstances of complicated tracking. This paper incorporated the contour characteristic of the objective into conventional methods and thus constructed a function to estimate the similarity of the objective. Experiments were conducted to show that more reliable results are obtained when rigid objects are tracked.