CAJ | 학술논문

为提高对复杂背景下形态剧烈变化目标的跟踪鲁棒性与跟踪精度，本文以粒子滤波为跟踪框架，改进并融合GrabCut目标分割技术。首先，采用J. Kwon等人提出的多特征、多帧混合模板作为粒子滤波的观测模型，构建粒子滤波跟踪框架。然后，改进GrabCut分割算法并提出多方向GrabCut目标分割算法，将其融入粒子滤波跟踪框架，进而提出本文的目标跟踪算法。实验结果表明，本文算法可对处于复杂背景下且形态剧烈变化的目标实现稳定跟踪，目标平均尺寸为70pixel×75pixel时跟踪平均误差为9.27 pixel。
위제고대복잡배경하형태극렬변화목표적근종로봉성여근종정도，본문이입자려파위근종광가，개진병융합GrabCut목표분할기술。수선，채용J. Kwon등인제출적다특정、다정혼합모판작위입자려파적관측모형，구건입자려파근종광가。연후，개진GrabCut분할산법병제출다방향GrabCut목표분할산법，장기융입입자려파근종광가，진이제출본문적목표근종산법。실험결과표명，본문산법가대처우복잡배경하차형태극렬변화적목표실현은정근종，목표평균척촌위70pixel×75pixel시근종평균오차위9.27 pixel。
To improve the robustness and precision when tracking target with fast changing appearance in complex background environment, particle filter was taken as tracking framework, GrabCut algorithm was improved and combined. First, multiple features and multiple templates composite model proposed by J.Kwon et al. was adopted as observation model of particle filter and particle filter tracking framework was designed. Then, GrabCut algorithm was improved and multi-directional GrabCut algorithm was proposed, and was combined in particle filter tracking framework, and this paper's tracking algorithm was proposed. The experiment results show that the proposed algorithm can track target with fast changing appearance in complex background environment stability, the tracking average error is 9.27 pixel when target average size is 75pixel×75pixel.