This paper describes a method which uses optical flow, that is, the apparent motion of the image brightness pattern in time-varying images, in order to detect and identify multiple motions. Homogeneous regions are found by analysing local linear approximations of optical flow over patches of the image plane, which determine a list of the possibly viewed motions, and, finally, by applying a technique of stochastic relaxation. The presented experiments on real images show that the method is usually able to identify regions which correspond to the different moving objects, is also rather insensitive to noise, and can tolerate large errors in the estimation of optical flow.