Segmentation Based On Human Motion Capture Data

In recent years, motion capture technology is widely used in many fields like movie animation, virtual reality, etc. The technology captures a performer’s trajectory according to a certain frequency. The motion information of each frame is saved to constitute the motion capture data. The current motion capture technology which has a high cost and needs professional actors makes the captured motion data longer. However, the long motion capture data makes data storage, retrieval, editing and compositing inconvenient. So the segmentation methods with high accuracy and high efficiency are required. The studies on segmentation methods have become an important research direction in related areasHuman motion segmentation based on capture data means segmenting long motion capture data sequence to get movement sections which contain different types of movement, so as to facilitate data storage, retrieval, editing and so on.In order to achieve human motion segmentation, firstly, the motion capture data in ASF/AMC format is parsed. The contents of each keyword in the files are described in detail. Then, according to the human skeleton model, the process of transforming capture data into continuous frame motion is presented. Finally, Euler angles and quaternions which can represent motion information are introduced, and the relationship between Euler angles and quaternions is described as well. The work lays the foundation for the feature extraction.In terms of segmentation algorithm research, first the segmentation method based on similarity search is studied. An improved method which converts the Euler angles data into quaternion logarithm feature data is proposed, and it is verified that its segmentation accuracy is higher than original algorithm by a large number of experiments.Then an improved human motion segmentation algorithm based on open-end dynamic time warping is implemented, a better result is obtained by quaternion logarithm feature extraction from the Euler angles data, as well as dimensionality reduction by Principal Component Analysis. The algorithm not only improves the segmentation accuracy, but also greatly improves the computational efficiency.Following that, the segmentation method based on generalized model with body partition index maps is introduced. The quaternion logarithm feature data is extracted from the Euler angles data and the restriction of the maximum length is introduced. Both of these make segmentation result better to some extent.Finally, a hierarchical aligned cluster analysis method for human motion segmentation is studied and improved. Motion data clustering is carried out to realize temporal reduction by mixture Gaussian models, and a self-adaptive method is presented for constructing the threshold parameter to compute the kernel distance. Experiment results demonstrate that the improved method gives a better segmentation.