The track-before-detection (TBD) based on a Rao-Blackwellized Particle Filter(RBPF) is proposed for detecting and tracking a number of weak targets in an infrared image with low Signal to Noise Ratio (SNR). A system dynamic model and a measurement model of the infrared image are presented. Then, the state vector is partitioned as a linear state variable and a nonlinear state variable with RBPF,which is computed by the Kalman filter and the particle filter,respectively. Finally,the present and absent statistics are defined and the flowchart of the RBPF-TBD for multi weak targets with a constrained condition are given. The simulation results of the infrared image sequences in different SNRs show that the RBPF-TBD can reduce the state’s estimation error and can improve the detection probability. For a single target by TBD,the simulation indicates that the consumed time is 0.0387 s/frame averagely. Under the same particle number,RBPF can get a highter filtering accuracy.For two targets with a SNR of 6 dB in different space positions, the simulation shows that constrained tracking can avoid the inference of present targets. The target presence and absence declaration delays averagely are 2.74 frames and 1.13 frames. These results show that the proposed algorithm for multi weak targets is capable of detecting the target presence and absence,and is capable of tracking the target efficiently.