在目标跟踪领域,常常通过建立先验模型,如路径一致性假设模型,对目标轨迹进行预测来处理跟踪过程中的遮挡问题.然而,当这种预测与目标的实际运动轨迹相差较大的时候就会发生跟踪失败.我们提出了一种交互式粒子滤波方法,通过判断不同目标样本观测之间的遮挡关系,自适应地选择不同外观模板进行相似性度量并更新粒子权值,成功地解决了跟踪过程中各目标之间的相互遮挡问题.实验结果表明,即使在目标间发生完全遮挡且被遮挡目标运动轨迹无法预测的时候,本算法仍然能够取得精确的跟踪结果.

When tracking multiple objects,prior models such as path consistency assumption model are generally established in order to handle occlusion problems.However,if the assumption is greatly distinguished from the real trajectories of the objects,the tracker is doomed to fail.To solve this problem,we put forward a definition of interactive particle filter which adaptively selects appearance template for a particle to measure its likelihood by judging the occluding relationship between each two samples of different objects.The experiments illustrate that our method accurately locates the object even if it is completely occluded as well as the trajectory is impossible to be predicted.