|Abstract:In order to guarantee the space-time adaptive processing performance of radar, it is necessary to ensure that training samples should be independent and identically distributed and share the same statistic property with the clutter in the cell under test (CUT). However, in the practical application, the training samples are likely to be contaminated by the inference target signals(outliers), resulting in the so-called target self-nulling phenomenon, and seriously degrading the detection performance. However, the conventional sample selection algorithms use the sample covariance matrix to represent the clutter characteristics of CUT. When the clutter characteristics of most samples deviate from the CUT, the representation is inaccurate, reducing the censoring efficiency. Therefore, this paper proposes a training sample selection algorithm for radar based on direct estimation of the CUT clutter characteristics. The proposed method directly uses the sub-aperture covariance matrix of CUT to characterize the clutter. Since the estimation process depends on no training samples, the estimation of CUT is not affected by the outliers. Moreover, considering the existence of target signal in the CUT, the proposed method removes the target component from the sub-aperture covariance matrix of CUT based on clutter covariance matrix reconstruction, which utilizes the clutter Capon spectrum integrated over a sector separated from the location of target. Compared with the traditional generalized inner product algorithm which uses single snapshot to calculate the detection parameters, the new algorithm uses the sub-aperture covariance matrix of the samples to characterize its statistical characteristics, obtaining more stable results. The simulation results show that the proposed algorithm selects training samples more accurately.