Labeled Multi-Bernoulli (LMB) filter was proposed based on classic Multi-Bernoulli filter by taking label space into account. It is a true trajectory filter in comparison with other methods. The motion model and measurement model of infrared small targets were studied. The LMB Track-Before-Detect algorithm was applied to detect and track infrared small targets. In order to detect and track multiple maneuvering targets with time-varying motion models, an IMM-LMB Track-Before-Detect algorithm was proposed by integrating the interacting multiple models (IMM) with LMB Track-Before-Detect filter. The sequential Monte Carlo (SMC) implementation for IMM-LMB track-before-detect was also given. Simulation results demonstrate that the proposed algorithm can detect and track multi-targets trajectories from raw image. In addition, it can update the model probability in an online way, which detects and tracks maneuvering targets adaptively.