Human motion capture (mocap) has gained an increasing attention in many applications, such as advanced human-machine interaction, computer animations, digital films, interactive games etc. A motion capture system is able to record and digitalize a sequence of human postures representing by the 3D coordinates of a set of body joints across time. However, as the motions are usually captured continuously, it is a time- and labor-consuming work to segment and label the data manually. On the other hand, in some collaborative applications, such as human-machine interaction, it is also required that the input motions should be learnt automatically and in real time, which in turn can drive the computer with corresponding react. Hence, efficient methods to recognize motion data are necessary. In this thesis, we develop new methods to address the mocap data recognition problem, which includes three sub-problems, i.e., isolated motion patterns recognition, sequential motion pattern recognition and real-time motion stream recognition. For isolated motion patterns recognition, the challenges are mainly caused by the high dimensionality and great variation of the data. Principle component analysis (PCA) is an efficient tool to reduce the dimension and extract features. However, it cannot retain the temporal information of the data points in samples when applied to time series data such as mocap data. Motivated by this, we propose two singular value decomposition (SVD) based methods named segmenal SVD (SegSVD) and bidirectional segmental SVD (Bi-SegSVD). They first segment the motion data into a certain number of sub-segments, and then process them with SVD in an accumulative manner along the forward direction for SegSVD and both forward and backward directions for Bi-SegSVD. Based on the segmental features, we calculate the similarity of two samples using a weighted dynamic time warping (DTW) based measure. The measure is further extended into a kernel function for support vector machine (SVM) classifier to classify the motion patterns. In sequential motion pattern recognition, an input motion is composed of multiple motion patterns with their categories and boundaries unknown in advance. Thus, an additional challenge, i.e., to detect the start and end points of the embedded patterns, is imposed on this task. To address this problem, two new approaches are proposed. First, we exploit an open-end DTW (OE-DTW) based scheme motivated by the fact that OEDTW is efficient in matching complete patterns with incomplete ones. By regarding the input motion as a complete pattern, and taking each of the template patterns as an incomplete pattern, we apply OE-DTW to find their optimal matched part, according to which, the embedded patterns are detected and recognized sequentially. Second, we take advantage of the SegSVD structure that is with multiple levels and detect the end points of the embedded patterns by referring to top levels of the template patterns using a new penalty based level matching scheme. In real-time motion stream recognition, it requires not only to identify and recognize the embedded patterns in input motions, but also to detect the unwanted motions, and the task should be finished in real time. Motivated by the fast speed and efficiency of the content-based indexing techniques, we introduce an body partition index map based approach for this problem. Noting that human motions are composed of the sub-motions of upper limbs, legs and torso, we partition the motions into five parts according to the five body partitions, and process the submotions with standard clustering techniques separately. A generalized model for each motion class is trained by integrating the projected cluster node strings of the training trials and five body partition index maps are constructed. During recognition, the input frames are projected into the clusters and then used to look up the index maps. With a flexible voting scheme and a set of end point detection conditions, the input motions are segmented and recognized as legal patterns or unwanted motions in real time. Finally, we apply the real-time recognition approach and develop an interactive dancing game system, in which users’ dance motions are lively captured and recognized, and according to the recognition result, the corresponding interactive motions are determined and used to drive the avatar’s animation. Hence, it provides an immersive environment that users can dance with the avatars (i.e. virtual partners) collaboratively.