CAJ | 학술논문

为了能更加准确鲁棒地跟踪目标，提出了特征加权融合的在线多示例学习跟踪算法（WFMIL）。WFMIL在多示例学习框架下分别训练两种特征（Hog和Haar）分类器。在跟踪过程中，通过线性运算融合成一个强分类器，同时在学习过程中对正包中的示例引入权重。实验结果统计表明WFMIL能很好地解决目标漂移问题，并且对目标遮挡、运动突变、光照变化以及运动模糊等具有较好的鲁棒性。
위료능경가준학로봉지근종목표，제출료특정가권융합적재선다시례학습근종산법（WFMIL）。WFMIL재다시례학습광가하분별훈련량충특정（Hog화Haar）분류기。재근종과정중，통과선성운산융합성일개강분류기，동시재학습과정중대정포중적시례인입권중。실험결과통계표명WFMIL능흔호지해결목표표이문제，병차대목표차당、운동돌변、광조변화이급운동모호등구유교호적로봉성。
For the object tracking problems in computer vision, feature Weighted Fusion online Multiple Instance Learning tracking algorithm(WFMIL)is proposed. WFMIL trains two features(Hog and Haar)classifier separately by multiple instance learning method. In the tracking process, they are integrated into a strong classifier by the linear operation. While in the learn-ing process, weight is introduced into instances of positive package. Experimental results show that WFMIL can solve the object drift and has a certain robustness in handling occlusion, target abrupt motion, illumination change, and motion blur.