CAJ | 학술논문

为了实现长时间稳定的对特定目标的跟踪，结合匹配型跟踪方法和决策型跟踪方法的优势，同时利用集成学习的思想构建多个强分类器，提出一种基于集成多示例学习的mean shift跟踪算法。首先在上一帧中对示例进行随机采样，构建分类器的集体，通过集成学习合成最终的分类器以确定当前帧中目标的初始位置；然后对初始位置和上一帧目标最终位置的距离与设定的阈值进行判断，决定是否采用mean shift跟踪算法对初始位置进行修订，以确定目标的最终位置。实验结果表明，该算法不但可以应对目标的形变、旋转、遮挡以及光照变化等各种复杂的情况，而且可以做到长时间的跟踪，具有较强的鲁棒性。
위료실현장시간은정적대특정목표적근종，결합필배형근종방법화결책형근종방법적우세，동시이용집성학습적사상구건다개강분류기，제출일충기우집성다시례학습적mean shift근종산법。수선재상일정중대시례진행수궤채양，구건분류기적집체，통과집성학습합성최종적분류기이학정당전정중목표적초시위치；연후대초시위치화상일정목표최종위치적거리여설정적역치진행판단，결정시부채용mean shift근종산법대초시위치진행수정，이학정목표적최종위치。실험결과표명，해산법불단가이응대목표적형변、선전、차당이급광조변화등각충복잡적정황，이차가이주도장시간적근종，구유교강적로봉성。
An effective object tracking method was proposed by combining multiple instance learning and mean shift tracking. The motivation is to use the advantages of the generative model and discriminative model and use ensemble learning to gain more robust tracking effect. First, instances are randomly selected to train different classifiers in the previous frame, and then the final integrated classifier was trained using ensemble learning to improve the tracking accuracy. The initial position of the object was determined by using ensemble multiple instance learning. Then mean shift tracking was used to revise the initial position by comparing the distance between the initial position and the object position in the previous frame to a threshold. The experimental results has shown that the proposed algorithm has good performance in many complicated situations, e.g. pose change, rotation, occlusion and changes of illumination, and can track successfully for a long time with strong robustness.