An improved dynamic programming track-before-detect (DP-TBD) algorithm is proposed in this paper. A new state transition probability model is developed based on the angle deviation between measurement and predicted state estimated by exponential smoothing. Then the transition probability is used to modify the scoring function of TBD. All these techniques are embedded in the recursion of DP to form a closed loop TBD system. Numerical simulations demonstrate that the proposed algorithm have better detection and tracking performance than the traditional one.