Regression PCA for Moving Objects Separation

This work proposed a new approach called regression PCA (RegPCA) for statistical machine learning and big data analyses. One of the potential use cases investigated in this work is to separate the moving objects (foreground) from the background images. This is achieved by performing regression before conducting Robust PCA (RPCA). RegPCA works well in the moving object detection task because the background information can be conceived as the regression portion of the images, while the residual portion of the regression can then be fed into RPCA to fine tune the foreground detection. The experiments show that in moving object detection problems RegPCA provides much better results than applying only RPCA, especially in color videos and when the moving objects are relatively big. Further studies are needed to leverage the interesting features of RegPCA approach and apply it to solve more real world problems.