Weighted residual minimization in PCA subspace for visual tracking

摘要: The success of sparse representation, in face recognition and visual tracking, has attracted much attention computer vision spite its computational complexity. These representation-based methods assume that the coding residual follows either Gaussian or Laplacian distribution, which may not be accurate enough to describe residuals real scenarios. In order deal with such issues a novel generative tracker is proposed Bayesian inference framework by exploiting both robust principle component analysis (PCA) algorithm. contrast existing algorithms, method introduces weighted least squares into PCA reconstruction avoiding complex l1-regularization. Further, it generate an occlusion map based on weights, used avoid updating information during incremental subspace learning. performance evaluation challenging image sequences demonstrates performs favorably when compared several state-of-the-art methods.