In this paper, a randomized method for detecting multiple ellipses based on the least square approach is presented. The main concept used is that we first randomly select three edge pixels in the image, which are the centre of three windows with the same size. In order to determine a possible ellipse, we use the least square method to fit all the edge points in these three window, and to solve the ellipse parameters through Lagrange multiplier method. Then we randomly select the fourth edge pixel in the image and define a distance criterion to determine whether there is a possible ellipse in the image. After finding a possible ellipse, we apply a further verification process to determine whether the possible ellipse is a true ellipse or not. Some artificial images with different levels of noises and some natural grey images containing circular objects with some occluded ellipses and missing edges have been taken to test the performance. Experimental results demonstrate that the proposed algorithm is faster and more accurate than other methods.