In order to detect and track low signal-to-noise ratio target, a new track before detect algorithm based on dynamic programming is proposed. The contributions consist of two aspects. Firstly, a new cost function method is proposed. It uses not only the measured amplitude information at current frame, but also uses amplitude association information between two consecutive frames. Secondly, an ergodic state backtracking algorithm with variable threshold is presented to detect new-born or vanished targets during the observation. Simulation results show the proposed algorithm effectiveness in detection performance.