Abstract: A compact form of the difference between two rotations represented by quaternion is derived, and the total rotation differences in all joints are taken as the distance between two frames. Based on this frame distance, an efficient algorithm is presented to extract key-frames from motion capture data. The whole animation process is reconstructed by simple linear interpolation of root joint translation and spherical linear interpolation of all joint rotations. Running experiment show that the original motion capture data can be compressed in a high ratio, the extracted key frames give a good visual summarization of the original animation and are useful for storing, transferring and browsing the motion data.