一种基于动态分层更新的多特征融合跟踪方法

摘要

针对单一特征的目标跟踪算法和传统的模型更新策略的不足,本文提出一种均值迁移和粒子滤波相结合的多特征融合跟踪方法.该方法通过均值迁移对粒子传播进行优化,根据粒子权值的分布情况自动调节各个特征的融合权值,实现了多特征的有效综合,通过建立目标模型的动态分层更新策略,有效保留了目标和场景的变化信息.实验结果表明,该方法对目标外观的快速变化适应性强,且实时性好,适用于动态场景下的跟踪.%Object tracking by using single feature results in a poor performance in complex scene, and the traditional template update strategy is not robust to target appearance changes. Therefore, the paper presents a multi-feature fusion tracking method by mean-shift and particle filter tracking framework. The particles were optimized by mean-shift thus overcoming the degeneracy problem. According to the distribution of particles, the fusion coefficient of each feature is estimated online. Besides, a dynamic layered-weighting strategy is used to adjust the updating weights adaptively according to the variety of each component in template. The experimental results show that our algorithm can deal with dynamic appearance changes of target, thus it is suitable for tracking in dynamic scene.