Human action recognition is a hot topic in computer vision. Most of the existing work use the action models based on supervised learning algorithms. To achieve good performance on recognition, a large amount of labeled samples are required to train the sophisticated action models. However, collecting labeled samples is labor-intensive. This paper presents a novel semi-supervised learning algorithm named multi-learner co-training model (MCM) to recognize human actions. Two key issues are addressed in this paper. Firstly, the base classifiers in co-training are selected by Q statistic-based classifiers selection algorithm (QSCSA). Secondly, MCM is employed for the semi-supervised model construction. The new confidence score measure of unlabeled sample depends on estimating the classifier companion committee (CCC) accuracy on labeling the neighborhood of an unlabeled examples. To evaluate the proposed algorithm, mixed-descriptors are used to express actions so that the recognition algorithm can quickly complete the recognition process from a single frame of visual image. Experimental results are presented to show that the proposed semi-supervised learning system can recognize simple human actions effectively.