基于偏最小二乘分析和稀疏表示的目标跟踪算法

经典的基于子空间学习的跟踪方法通过主成分分析(principal component analysis, PCA)建立并更新目标的特征模型，只考虑目标的特征信息而忽视模型中每个样本的类别特征，从而降低了目标的跟踪精度。为此，提出一种基于偏最小二乘分析(partial least squares analysis, PLS)和稀疏表示的目标跟踪算法。通过PLS去分析关于目标与背景中纹理特征和类别信息之间的相关性，建立一个可区分的低维特征空间。将目标模板线性表示为所有目标候选的线性组合，当存在与目标模板相似的候选时，线性表示的系数满足稀疏性约束，通过L1范数最优化求解稀疏表示系数，根据最小的重构误差得到最优的目标位置。在多个视频场景下的实验结果展示了改进的跟踪算法具有较好的跟踪性能。

Object tracking via partial least squares analysis and sparse representation

The classical object tracking methods based on subspace learning build and update object''s feature model by principal component analysis (PCA), and they only use feature information and their formulations do not exploit the label information to reduce the precision of object tracking. An improved object tracking method based on partial least squares analysis and sparse representation is proposed in this paper. The correlation of object appearance and class labeled information from foreground and background is modeled by partial least squares analysis, for generating a low-dimensional discriminative feature subspace. The target template is represented by a set of all target candidates. Assuming that there is one target candidate similar to the target template, the representation coefficients will be sparse, which can be obtained via L1-norm minimization. Optimal target position is obtained by the smallest reconstruction error. Experimental results demonstrate favorable performance of the proposed tracking algorithm.