CAJ | 학술논문

为了解决Mean Shift跟踪算法中目标模板只能从单一图像建立且很难更新问题，提出了一种结合改进的Mean Shift与增量式支持向量机的红外目标跟踪算法。首先，根据目标区域的灰度直方图对目标进行描述，然后采用标准Mean Shift搜索目标，结合子图图像矩特征进行二次搜索，再计算下一帧搜索的窗口大小，以解决目标尺寸明显变化时造成目标丢失的问题。同时，针对目标遮挡易导致跟踪失败的问题，引入机器学习理论，采用增量式支持向量机自适应更新模板，则目标跟踪问题转换为目标和背景的分类问题。实验结果表明：提出的改进算法在目标尺寸、姿态发生变化或出现部分遮挡时，能有效跟踪目标。
위료해결Mean Shift근종산법중목표모판지능종단일도상건립차흔난경신문제，제출료일충결합개진적Mean Shift여증량식지지향량궤적홍외목표근종산법。수선，근거목표구역적회도직방도대목표진행묘술，연후채용표준Mean Shift수색목표，결합자도도상구특정진행이차수색，재계산하일정수색적창구대소，이해결목표척촌명현변화시조성목표주실적문제。동시，침대목표차당역도치근종실패적문제，인입궤기학습이론，채용증량식지지향량궤자괄응경신모판，칙목표근종문제전환위목표화배경적분류문제。실험결과표명：제출적개진산법재목표척촌、자태발생변화혹출현부분차당시，능유효근종목표。
In order to solve the problem that the target template of standard Mean Shift tracking can only be built from a single image, and difficult to update, an algorithm combining improved Mean Shift with incremental Support Vector Machine for infrared target tracking was proposed. First, target was described using gray histogram of the target region. Then, in order to solve the problem of target lost in tracking caused by target size obviously changing, target localization was started using standard Mean Shift, and then image moment feature of the sub image for secondary search was combined to calculate the tracking window size for next frame. Meanwhile, according to the problem of target occlusion easily lead to tracking failure, machine learning theory was introduced and incremental support vector machine was used to update target template adaptively, thus target tracking problem was converted to a problem of classification between the target and the background. Experiments show that the improved algorithm proposed in this paper performs well even if greatly change occurs in target pose, size or partial occlusion happens.