鲁棒的交互式多模粒子滤波跟踪器概率计算方法

针对传统交互式多模粒子滤波(IMMPF)跟踪器概率 计算方法在复杂环境下鲁棒性不足的缺陷,提出 一种新型的基于联合似然函数模型的子跟踪器概率计算方法。首先,计算基于跟踪结果与当 前外观 模型的巴氏距离作为瞬时似然函数,度量目标外观的剧烈变化；其次,利用l ₂范数规则化最小二乘算法构 建目标的重构外观模型,将其与跟踪结果的误差指数函数作为平稳似然函数,度量目标的缓 慢变化； 然后,基于加权求和策略得到跟踪器基于多种特征的联合似然函数；最后,将建立的联合似 然函数结 合上一帧的先验状态交互概率完成子跟踪器概率的更新。对复杂环境下跟踪器性能的在线评 估对比 结果验证了联合似然函数模型能有效评估跟踪器因不同干扰因素导致的性能变化,将其应用 于子跟踪器概率的计算能获得比主流算法更好的鲁棒性。

Robust interactive multiple model particle filter tracker probability calculation method

As the traditional interactive multiple model particle filter tracker prob ability calculation method has poor robustness in complex environment,a novel sub-tracker probability calculation method based on collaborative likelihood function is proposed.Firstly,the Bhattacharyya distance between tr acking result and immediate appearance model is measured as the instantaneous likelihood function to evalua te abrupt appearance changes, and the error exponential function between tracking result with reconstruction a ppearance model obtained from l2regularized least sequel algorithm is measured as the stable likelihood funct ion to evaluate slow changes. Then,the collaborative likelihood function based on multiple features is measur ed by weighted sum strategy. Last,tracker probability is updated by combining collaborative likelihood funct ion and prior state interacting model.Comparative results of online evaluation for tracker quality in complex e nvironment demonstrate that the collaborative likelihood function can effectivel y evaluate the changes of tracker quality caused by different interference factors,and utilizing it to calculate tracker probability can get better robustness than state-of-the-art algorithms.