粒子滤波跟踪中反馈式多模板更新策略研究

Feedback-Loop Multi-Template Updating Algorithm in Particle Filter Tracking

在粒子滤波跟踪算法中,目标模板的更新精度是影响目标跟踪稳定性的重要因素.传统的模板更新策略大多以固定帧频或简易的更新判定等方式选取目标作为模板训练的样本,难以解决遮挡和目标外观突变等问题,并且不具备从错误模板中恢复的能力.针对该问题,在增量子空间框架下提出了反馈式多模板更新策略.首先引入稳态目标模板、暂态目标模板以及背景模板建立多模板模型,然后利用候选目标与多模板模型的相似度等信息,准确判定当前的跟踪状态及干扰产生的原因,并将其作为模板更新时机和方法的选择依据,形成闭环反馈的模板更新机制.实验结果表明,该方法能够有效克服遮挡和光照等复杂环境的影响,又能较好地适应目标外观的突然变化,比传统模板更新算法具有更好的鲁棒性.

In particle filter tracking,the accuracy of template is essential to the tracking stability. Traditional updating strategy of template usually choose current target as the sample for training new template at a constant frequency or by simple judgment,which makes it impossible track target robustly under occlusion and abrupt target appearance changing,additionally hard to reject failure template for better one.To solve these problems,a feedback-loop multi-template updating algorithm is proposed based on incremental subspace learning.Firstly,steady target template, transient target template and background template are introduced to establish multi-template model.Secondly,those information such as the similarities of candidate target and multi-template, are applied to judge the present tracking state and the cause of interference,which are to choose the updating method and timing.Thus,a feedback-loop updating mechanism of target template is built.Experimental results indicate that such an updating strategy can overcome the influence of complex environment such as occlusion and illumination,as well as be adaptive to the appearance changes and is more robust than the traditional template updating strategy.