CAJ | 학술논문

由于光照的变化、形变、姿势的变化，传统的粒子滤波算法会发生跟踪漂移。影响了跟踪算法的准确性和稳定性。发生漂移的主要原因是目标函数不够平滑，从而使得更新的状态存在较大误差。本文提出了一种新型的层叠式粒子滤波算法，该方法在离散域提取基于核的直方图作为目标模型，通过改进传统的粒子滤波算法，实现对目标由粗到细的跟踪。在公开测试集上与当前主流的跟踪算法比较，结果表明本文提出的算法在跟踪精度上提高了20%，并取得了很好的实时性和鲁棒性。
유우광조적변화、형변、자세적변화，전통적입자려파산법회발생근종표이。영향료근종산법적준학성화은정성。발생표이적주요원인시목표함수불구평활，종이사득경신적상태존재교대오차。본문제출료일충신형적층첩식입자려파산법，해방법재리산역제취기우핵적직방도작위목표모형，통과개진전통적입자려파산법，실현대목표유조도세적근종。재공개측시집상여당전주류적근종산법비교，결과표명본문제출적산법재근종정도상제고료20%，병취득료흔호적실시성화로봉성。
In the framework of traditional particle filter, object is always drifted for theillumination light changing, deformation and poses changing, which affects the accuracy and robust performance of tracker. The reason is that there is not smooth in the objective function. Error will accumulate during the state updating process. A new stack particle filter is proposed to represent target model by the kernel histogram in the discrete fields. It can track the target from coarse to fine by the modified particle filter. Extensive experimental results show that the proposed algorithm performs 20% improvement in the accuracy, favorably time expedition and robustness against state-of-the-art methods.