Automatic classification of vault jumps using video analysis
P.L. Oppedijk · Research Repository (Delft University of Technology) · 2013
In sports, the use of motion-capture techniques increases, leading to a fast increase in valuable motion data. Automatic recognition and classification of the captured motions, provides an orderly structuring of the motion data. By this the users can easily retrieve specific motion data. In this thesis, we consider the automatic classification of vault jumps in gymnastics, captured by a high speed video camera system. A vault jump consists of a sequence of motions belonging to a predefined motion label, such as a Handspring. Then, the vault classification problem consists of automatically recognizing a vault jump in a video recording and assigning the appropriate label to the recording. To this end, we segment the vault classification problem into a sequence of vault-section classification problems. The following vault-sections are proposed; Type of Vault (TV), Number of Somersaults (NS), Type of Somersault (TS) and Number of Twists (NT). The segmentation into vault-sections allows for the development of a versatile classification system, capable of classifying a large number of vault classes based on a limited amount of data. Next, we use video analysis techniques to transform a video recording into feature representations, or so called feature sets, which reflect the specific characteristics of the vault jump throughout the four vault-sections. The four vault-section feature sets are then classified, resulting in four vault-section classifications. The final labeling of the recording of a vault jump is by the combined results of the four vault-section classifications. The proposed automatic vault classification system is based on the vault jump recordings made by Van de Eb et al. [1] at the world championships in gymnastics 2010. Extensive experiments have been conducted on these recordings for evaluating various feature sets and classifiers per vault-section, resulting in one best performing combination per vault-section. Furthermore, the vault-section classifications are evaluated on their influence on the classification performance of a complete vault jump. In the end, an overall classification rate of 69.5%, with a correct classification accuracy of 90.2%, is obtained for the classification of the vault jump recordings.