In this paper we consider the problem of tracking multiple targets. Following a mean field approach, we obtain a cost function that depends on the means and covariances of each object. A simplification of this cost function leads to self consistent equations for the means, while the covariances are obtained using a normal Kalman filter. The iteration of the self consistent equations allows refining the solution and including the effect of occlusions. An implicit assumption of our model is that we can deduce the depth order from the joint state and that we can approximate the posterior of each object by a Gaussian. We show how this simple approach reduces the number of tracking failures in two sequences. The first one is a synthetic sequence of tennis balls moving on a background. The second one is a sequence of a football match taken from the VS-PETS data base.