CAJ | 학술논문

为了增强基于分布场跟踪算法在复杂场景下的性能，提出了一种基于局部分布场的目标跟踪算法。首先将整体目标分布场模型随机分成多个局部分布场；然后基于前景分布场与背景分布场的 L1范式距离越大，越能将前景从背景中区分出来的原则，选择若干个具有较好判别性能的局部分布场；最后基于所选择的局部分布场，利用梯度下降法搜索目标。实验结果表明，在复杂场景下，该算法比当前流行的分布场跟踪算法和多示例学习跟踪算法更加鲁棒。
위료증강기우분포장근종산법재복잡장경하적성능，제출료일충기우국부분포장적목표근종산법。수선장정체목표분포장모형수궤분성다개국부분포장；연후기우전경분포장여배경분포장적 L1범식거리월대，월능장전경종배경중구분출래적원칙，선택약간개구유교호판별성능적국부분포장；최후기우소선택적국부분포장，이용제도하강법수색목표。실험결과표명，재복잡장경하，해산법비당전류행적분포장근종산법화다시례학습근종산법경가로봉。
In order to enhance the performance of distribution fields (DF) based tracking algorithm in complex scenarios ,a local distribution field based tracking method is proposed in this paper .Firstly , we randomly crop the whole DF into multiple local DFs .Secondly ,based on the principle that larger L1 distance between foreground DF and background DF can better distinguish foreground from background ,some discriminative local DFs are selected .Thirdly ,we search the target by a simple gradient descent based on those selected local DFs .Experiment results show that the proposed local DFs method performs better in complex scenarios than the popular DF and multiple instance learning (MIL) tracking algorithms .