In this paper, we present a multiple target tracking (MTT) algorithm for time-varying number of targets with linear state dynamics and a non-linear observation model. The algorithm uses a particle filter to target the joint posterior of data association and target states given observations. Target states are inferred by a Rao-Blackwellised particle filter which integrates out the velocity part of a target state, leaving only its position part to be sampled. We also design an efficient Markov chain Monte Carlo (MCMC) kernel to rejuvenate target positions in the spirit of the resample-move algorithm. Simulation results show that Rao-Balckwellisation of the velocity component and the additional MCMC move lead to a notable improvement over the standard particle filter for MTT.