The traditional track-before-detect algorithms based on non-coherent integration demand high SNR for each single frame data, and the effective time for signal integration is short. In this paper, a novel and efficient track-before-detect algorithm is proposed based on the particle filter with coherent integration. The target states are estimated by the samples of the delay, velocity, and power information of target. Then, the matched filters are redesigned according to the target delay, and the echoes are filtered again to compensate for their delay. The simulation results show that: at the low SNR of -5 dB, the algorithm still has fine tracking and signal detection performances by extending time for signal integration.