Key motion spotting in continuous motion sequences using motion sensing devices

This paper proposes a key motion spotting method which is designed to locate the motion of interest (or key motion) from a database of motion sequences. As a key motion could be just a subsequence of the stored motion sequence, the proposed method differs much from the general methods of content-based retrieval of motion sequences, which searches from the database the individual sequences similar to a query motion sequence. The key problem addressed here is how to locate the points where a subsequence similar to the key motion starts and ends inside each database sequence. Aiming at this problem, we propose an ending/starting point detection method based on the Dynamic Time Warping (DTW). To evaluate the performance of the proposed method, we have implemented the method in an experimental motion query system which can capture motion sequences with the Microsoft Kinect. To characterize the body motion, two different kinds of features, called the motion histogram feature and the limb vector angle feature, are extracted from the acquired motion data. From the experimental results, the system achieves an average precision rate of 63% and an average recall rate of 70%. The speed of matching is 33.3 frames per second.