CAJ | 학술논문

提出一种多特征融合和带宽自适应均值偏移跟踪算法。基于M-估计器建立位置和带宽关于完整参数型带宽矩阵的迭代更新公式。分析权值图像的本质，基于灰度和局部标准差建立融合权值图像。通过先前目标模板模型和确定的目标模型的平均值生成当前目标模板模型。在位置向量的迭代公式中，采用扩大的带宽矩阵，确保定位精度。为防止由于背景杂波导致带宽膨胀或者由于目标自我相似导致带宽收缩，引入规范准则。跟踪的视觉结果和评估尺度表明，提出的跟踪算法相比于另外三种尺度自适应均值偏跟踪算法，具有最好的性能。
제출일충다특정융합화대관자괄응균치편이근종산법。기우M-고계기건립위치화대관관우완정삼수형대관구진적질대경신공식。분석권치도상적본질，기우회도화국부표준차건립융합권치도상。통과선전목표모판모형화학정적목표모형적평균치생성당전목표모판모형。재위치향량적질대공식중，채용확대적대관구진，학보정위정도。위방지유우배경잡파도치대관팽창혹자유우목표자아상사도치대관수축，인입규범준칙。근종적시각결과화평고척도표명，제출적근종산법상비우령외삼충척도자괄응균치편근종산법，구유최호적성능。
A multiple features fusion and bandwidth adaptive mean shift tracking algorithm had been proposed. The iterative solution expressions of location and bandwidth had been established with fully parameterized bandwidth matrix based on M- estimator. The fused weight image had been produced with intensity and local standard deviation. The target template model had been generated by the combination of previous target template model and the mean value of the determined target models in the previous frames. An enlarged bandwidth matrix had been employed in the iterative solution of location vector to ensure location accuracy. To prevent the bandwidth from exploding in the presence of background clutter or imploding on self- similar target, regularization terms had been introduced. The visual results and evaluation measures show that the proposed tracking algorithm has the best performance compared with other three scale adaptive mean shift tracking algorithms.