摘要
利用单一特征在复杂环境下进行目标跟踪容易导致跟踪失败。针对该问题,提出基于多特征融合与均值偏移的粒子滤波跟踪算法。在粒子滤波的总体框架下,通过嵌入均值漂移聚类算法产生更逼近真实后验分布的粒子,同时采用颜色和结构特征作为观测模型来表示目标,利用融合后的信息计算粒子的权值,并在跟踪过程中不断更新,以减小跟踪偏差。实验结果表明,与基于颜色与结构的跟踪算法相比,该算法在使用相同粒子数目时鲁棒性更高,而且粒子的平均权重得到了提高,重采样次数明显减少,即使在粒子数目较少的情况下也能实现稳定跟踪。
To solve the problem that a single feature leads to tracking failure easily in a complex environment,a Particle Filtering（ PF） tracking algorithm based on multi-feature fusion and Mean Shift（ MS） is proposed. Under the framework of PF,it is closer to the real posterior distribution by embedding MS algorithm and using color and structural as the observation model to represent the object,and the weights of particles are calculated by this integration,in order to reduce the tracking deviation. Experimental results show that the proposed algorithm has better robustness when using the same particles,and the average weight of the particle is improved and the resample times are reduced significantly,even using the less particles can achieve tracking stability.
出处
《计算机工程》
CAS
CSCD
2014年第11期14-17,共4页
Computer Engineering
基金
河南省重点科技攻关计划基金资助项目(122102310309)
河南省基础与前沿技术研究基金资助项目(142300410147)
河南理工大学博士基金资助项目(B2011-58)
关键词
目标跟踪
均值偏移
多特征融合
粒子滤波
颜色特征
结构特征
object tracking
Mean Shift （ MS ）
Particle Filtering （ PF ）
multi-feature fusion
color feature
structural feature
