Abstract: When the moving object is occluded, or similar to background, it is hard to track the moving object. An optimize particle filter arithmetic based on adaptive features fusion mean shift method is proposed to solve this problem. This paper uses united histogram to describe the grayscale and gradient direction features of the object, adjusts features weight adaptively based on the features dependability of the object of prior picture. In particle filter theory it uses the improved mean-shift method to make particles of the particle filter to move towards estimated direction of maximal posterior kernel density of the target state, and designs a fusion observational model to improve the scene adaptability. The experimental result show that this algorithm can track military camouflage target reposefully where the color for both target and background are similar, and is robust for serious occlusion.

Key words: object tracking, features fusion, mean-shift, particle filter

摘要： 针对军事伪装目标在运动过程中存在与背景分布十分相似或遮挡等强干扰情况下的跟踪问题，提出了一种基于自适应多特征融合的均值漂移算法优化的粒子滤波跟踪算法。利用背景加权后的联合直方图表述目标灰度和梯度方向信息，根据前一帧目标特征的可信度自动调节双方的权重，在粒子滤波算法的框架下，利用改进后的均值漂移算法使粒子向目标状态的最大后验核密度估计方向移动，并设计了特征融合的观测模型，以提高跟踪算法的场景适应能力。实验结果表明，该算法可实现对与背景相似的军事伪装目标的稳定跟踪，对目标的严重遮挡具有很好的鲁棒性。

关键词: 目标跟踪, 特征融合, 均值漂移, 粒子滤波