In order to solve the problem of high probability of misclassification for similar samples of the collaborative representation classifier (CRC), we propose a discriminative CRC (DCRC), which takes the effect of all training samples and each class of training samples on the collaborative representation coefficient into account. The coefficient obtained has strong discrimination and can improve the discriminability of the similar samples. Human action recognition is conducted based on DCRC. We first extract the features of depth action sequence via depth motion maps (DMMs). Then, we use DCRC to encode the DMMs features and perform classification and recognition by new classification rules. Experimental results on the human action recognition datasets show that the DCRC has certain discriminative properties for similar actions, and the recognition accuracy is superior to some existed methods.