基于双模型融合的自适应目标跟踪算法

Adaptive target tracking algorithm based on fusion of two models

针对目标跟踪过程中的光照变化、背景混乱和目标形变等问题,提出一种背景抑制的HS直方图和核相关滤波双模型融合的自适应跟踪算法。首先引入非线性核相关滤波跟踪模型;其次提出背景抑制的HS颜色直方图跟踪模型,通过分离亮度分量以减小光照干扰,并采用背景加权突出目标信息;然后提出一种自适应融合策略,根据目标与背景的HS特征相似度来动态调整两个模型融合权重,以降低背景混乱和目标姿态变化的影响;最后针对目标尺度变化问题,采用尺度金字塔估计策略进行解决。在多个公开数据集下的对比实验表明,与现有算法相比,提出的算法能更好地降低光照、背景混乱等复杂因素干扰,且达到了工程应用的实时性要求。

To deal with illumination variation,background clutters and object deformation,this paper proposed an adaptive tracking algorithm using the fusion of background suppressed HS histogram model and correlation filter model. Firstly,it introduced a non-linear kernelized correlation filter tracking model. Secondly,it proposed a background suppressed HS histogram tracking model. This model separated luminance component to reduce illumination interference and used a background weighted method to highlight object information. Furthermore,this paper proposed an adaptive fusion strategy. According to the HS similarity between object and each background patch,it adjusted the fusion weight of two models dynamically to reduce the influence of background clusters and pose variation. Finally,it used a scale pyramid estimation strategy to handle scale variations. Experimental results on public datasets demonstrate that,compared with other trackers,this algorithm performs better on complex factors,such as illumination variation and background clusters,and meets engineering application real-time requirements.