We present a method to extract foreground object regions efficiently from image sequences. Scale-invariant feature transform algorithm is adopted to estimate the descriptor firstly by matching between two consecutive frames. Given local descriptor matching results, dense motion vector of each pixel is calculated by large displacement optical flow with variational optimization, which integrates detailed descriptors into the variational model. Then the foreground object boundaries and regions are detected by computing the optical flow gradient and magnitude. Experiments demonstrate that the method can achieve better segmentation results than alternative methods and adapts well to moving objects in relatively stationary background image sequences.