Track-before-detect (TBD) technology based on the probability hypothesis density (PHD) filter can effectively solve the problem of tracking number-varying dim multi-target. The main limitation of the standard PHD-TBD algorithm is the estimation error of target numbers influenced by the measurement noise remarkably. An improved PHD-TBD algorithm based on the smooth is proposed. The algorithm can overcome the influence of noise in a certain extent by updating the weight of particle using forward recursion and backward smooth, and then a steady estimation of target numbers is obtained. In addition, the simulation results demonstrate that the proposed algorithm can effectively and stably estimate the number of targets and their positions comparing with the standard one.