CAJ | 학술논문

针对传统的基于纹理特征的跟踪算法采用统计直方图描述目标，表达不够精确且计算量较大的问题，提出一种基于多级量化的局部纹理特征的点对点的图像跟踪算法。首先定义了一种高维纹理特征，其对单个的点具有较强的描述能力；随后通过矢量运算进行点对点匹配，根据匹配得出的点集的空间位置，利用均值漂移矢量实现初步定位；最后通过连续缩放窗口比例，同时使用颜色信息验证候选窗口与原始窗口的相似度，确定目标的位置和尺度。实验结果表明，该算法在场景复杂、噪声较大的环境中仍能准确地搜索到目标；在目标尺度变化时也具有较强的鲁棒性。
침대전통적기우문리특정적근종산법채용통계직방도묘술목표，표체불구정학차계산량교대적문제，제출일충기우다급양화적국부문리특정적점대점적도상근종산법。수선정의료일충고유문리특정，기대단개적점구유교강적묘술능력；수후통과시량운산진행점대점필배，근거필배득출적점집적공간위치，이용균치표이시량실현초보정위；최후통과련속축방창구비례，동시사용안색신식험증후선창구여원시창구적상사도，학정목표적위치화척도。실험결과표명，해산법재장경복잡、조성교대적배경중잉능준학지수색도목표；재목표척도변화시야구유교강적로봉성。
To solve the problem of inaccuracy and large calculation of traditional histogram-based tracker, this paper proposes a local feature-based tracking algorithm using multi-scale texture descriptor for each point. A high-dimensional texture descriptor with discriminative power is defined. Vector operation is applied to points matching and primary mapping is implemented through mean-shift according to the position of spatial point set. Subsequently, the location and scale of the object is determined by verifying the similarity between candidates and original target according to color information while the window is zooming. The results show the robustness of this algorithm against clutter, noise and change in target scale.