基于多个颜色分布模型的粒子滤波跟踪算法

Particle filter tracker based on multiple color distribution models

基于粒子滤波的目标跟踪性能在很大程度上依赖于观测模型的选择。为了解决被跟踪目标外观特征变化导致模型漂移问题,提出了一种新的粒子滤波算法,利用目标外观的先验知识,为目标建立多个颜色模型,通过对目标函数的优化,采用最优凸组合模型实时地对目标进行跟踪。同时,采用UKF(Unscented Kalman Filter)产生粒子滤波的建议分布,并从中抽取粒子。由于考虑到当前观测,使得粒子分布更加接近后验概率分布,用较少的粒子就可以逼近目标的真实状态。实验结果表明,与单一模型以及自适应模型算法相比,多模型的粒子滤波算法能够有效处理由于目标外观变化而导致跟踪性能下降甚至丢失目标的问题,而且计算代价不大。

The performance of target tracking via particle filtering are closely related to the observation model. A novel particle filter algorithm has been proposed to solve the model shift caused by appearance changes of the target. We model the multiple color distributions according to prior knowledge of the target. Through minimizing the cost function, the optimal convex combination model is selected to track the target in real time. Meanwhile utilizing UKF（Unscented Kalman Filer） generates the proposed particle distribution and extracts particles. Due to considering the current measurements, the distribution of particles is closer to posterior probability density, and a smaller number of particles can be approximated to the true state of the target. Experiment results show that the proposed algorithm has better tracking precision than the standard particle filter and low computational cost.