Different from the existing researches on target tracking using a fixed motion pattern, an algorithm is proposed for online recognition of human motion patterns with a monocular camera, to develop a human tracking method. First, moving target is detected using visual information and the visual features are extracted. Then, the motion features of target are acquired by depth extraction algorithm of monocular vision. The differences of a couple of successive frames' features are fed into random forest (RF) classifier to recognize human motion patterns online. Finally, different target motion models are chosen online based on classification results and approximate optimal particle filter is used to realize accurate estimation of human’s states. Extensive experimental results demonstrate the effectiveness of the proposed algorithm.