融合模糊统计纹理特征的多线索粒子滤波跟踪

Multi-cue particle filter tracking based on fuzzy statistical texture features

针对粒子滤波跟踪算法使用单一特征鲁棒性差,以及粒子重采样策略易导致粒子退化、贫化等问题,提出了一种基于多特征、多线索的改进粒子滤波跟踪方法。使用引入邻域关系的Histon直方图描述目标的颜色特征,并建立了一种稳健的模糊统计纹理特征(FSTF)表达空间纹理信息,然后将其自适应地与区域颜色特征融合构建多线索的观测模型。在粒子滤波跟踪过程中,使用基于K-means的粒子权重聚类进行更为准确的后验分布估计。在重要性重采样阶段,保留高权重粒子的同时基于当前目标状态的先验分布产生新粒子,避免了粒子退化并保证了粒子的多样性。在标准测试集上的实验结果表明:相比其他基于粒子滤波框架的跟踪算法,本文方法能够得到更高的跟踪精度和成功率。与其他效果突出的流行跟踪算法相比,本文方法能在光照变化、目标形变和背景扰动场景下取得更好的跟踪效果。

In order to solve the problem that particle filter tracking algorithm uses single feature with lower robustness and the resample strategy is easy to cause particle degradation and impoverishment,an improved particle filter tracking method based on multi-feature and multi-cue is proposed.First,the Histon histogram that introduces the neighborhood relationship is used to describe the color characteristics of the target.Then a robust fuzzy statistical texture feature is used to express spatial texture information,and it is adaptively fused with regional color features to construct a multi-cue observation model.In the particle filter tracking process,K-means-based particle weight clustering is used for more accurate posterior distribution estimation.In the importance resample stage,high-weight particles are retained while new particles are generated based on the prior distribution of the current target state,which ensures particle diversity and avoids particle degradation.Experiments are carried out on the standard test set.Compared with other tracking algorithms based on particle filter framework,the proposed method obtains higher tracking accuracy and success rate.Compared with other popular tracking algorithms,the proposed method can achieve better tracking results under illumination variation,target deformation and background disturbance scenarios.