A new recognition method is developed to relax the heavy requirement of training sample size in the radar high resolution range profile (HRRP) target recognition. The statistical characteristics of HRRP′s frequency spectrum amplitude (FSA) are analyzed. Then a linear Gaussian mixture state space model (LGMSSM) is proposed to describe the stationarity and multimodality of the FSA. Afterwards, an expectation maximization (EM) algorithm is derived for model parameter estimation. Finally, experimental results based on measured data show that the proposed method can obtain satisfactory recognition accuracy and rejection performance even with a very few training samples.