Radar target detection in complicated clutter environments is always a challenging task. Statistical model-based detectors are restricted to the assumed radar clutter distribution and feature-based detectors rely on handcrafted features, leading to serious degradation in the face of clutter with complex spatio-temporal variations. To overcome the above limitations, in this paper, we introduce metric learning into radar target detection to learn metric matrices directly from the source data for classification. First, based on the large margin nearest neighbor (LMNN) and kernel support vector machine (K-SVM), a classification framework called LMNN-KSVM is provided. The framework integrates the LMNN and K-SVM using the Mahalanobis distance-based Gaussian kernel function. Subsequently, a detector based on the LMNN-KSVM is then developed. A data augmentation method for the Doppler amplitude spectrum of target samples is also presented. Second, an online version of the LMNN-KSVM-based detector is further developed by considering the limited data condition of the practical radar detection scenario (i.e., only a short front segment data of the slow-time domain is available for training). A quick update method for the Mahalanobis matrix is devised and the following testing data are utilized to continuously update the detector, enabling the detector to adapt to the temporal fluctuation of the target and radar clutter. The proposed detectors are applied to sea-surface target detection and experiments are conducted on the IPIX database, which demonstrates superior performance of the proposed detectors.