The technique for moving object detection consists of two major phases:
Image filtering of a current frame with the noise removal filter coefficients extracted with the PCA analysis.
Detection of moving objects applying the pixel based method for moving object detection resistant to illumination changes.
A given video is treated as three-dimensional (3D) array of gray pixels with two spatial dimensions X, Y and one temporal dimension Z. We use spatiotemporal (3D) blocks represented by N-dimensional vectors, where a block spans (2T+1) frames and contains NBLOCK pixels in each spatial direction per frame N = (2T+1)× NBLOCK× NBLOCK, which can be found in the literature mentioned in the references. To represent the block vector by a scalar while preserving information to the maximal possible extent, principal component analysis is used. For principal component analysis, sample mean and covariance matrix of representative sample of block vectors corresponding to the considered types of movies are estimated and the first eigenvector of the covariance matrix S (corresponding to the largest eigenvalue) is used that represents the coefficients of the 3D filter that suppresses the noise. The 3D filter can be emulated by three 2D filters applied on frames z-1, z and z+1.