传统多通道跟踪常采用少数离散通道并认为通道间互不相关, 忽略了多通道特征渐变的“多峰”型可信度分布。本文提出一种基于多通道能量极值点集的两层粒子滤波目标跟踪方法。首先将目标分解为多个频率、多个角度的通道, 提取极值点集合作为目标特征。其后提出两层粒子滤波框架, 以一级粒子作为目标运动状态整体性约束; 二级粒子表征其下属的多通道特征, 通过局部优化方法使其聚集在特征峰值附近。以多通道特征的可信度对二级粒子进行加权, 并反馈至一级粒子作为其加权值, 兼顾了目标整体约束以及对多特征可信度分布的表征。跟踪仿真实验结果表明, 该算法计算量适中, 工作稳定, 与对比算法比较, 平均跟踪误差减小了约 0.5个像素。

Classical multi-channel algorithms usually employ minority discrete channels. It is considered that the channels are independent and their multi-mountain like distribution is neglected. Double-layer particle filter for target tracking with multi-channel energy extreme points set method is proposed. Visual target is firstly decomposed to multiple frequency and angle channels. And then energy extreme points set is calculated as target feature. Double-layer particle filter algorithm framework is established. Target motion state is globally constrained by first-order particle. Second-order particles represent the accessory multi-channel features and cluster to feature peak by local optimization method. Multi-channel features are examined and then second-order particles are weighted. The result is feedback to first-order particles as weighting parameters. Thus global constraint and multi-feature reliability distribution are balanced. Experimental results show that the algorithm runs steadily with moderate computational complexity. Compared to correlation methods, the average tracking error is decreased by about 0.5 pixels.