Automatic Classification System for Vault (ACSV):
To efficiently handle large amounts of video data in gymnastics, Rick Oppedijk developed, in collaboration with InnoSportLab ’s-Hertogenbosch (ISL-DB), the University Amsterdam (VU) and the Technical University Delft (TU Delft), an automatic classification system based on video analysis which is capable of automatic recognition and classification of vault recordings.
The concept of an automatic classification system for vault originated after a pilot study in inertial sensors as information source for gymnastics. Three bachelor-students of the TU Delft conducted the study at ISL-DB. The study resulted in a concept classification system based on inertial data capable of classifying four different vault classes.
The initial approach of the thesis study of Rick Oppedijk was to expand the concept system into a prototype, which in theory would be capable of classifying every known vault class and by this act as a Proof of Concept. The following requirements where set:
- Proof of versatility, such that in theory every known vault class can be classified.
- Proof of sensitivity, to separate vaults that are in a biophysical way much alike.
- Proof of robustness, to be able to classify similar vaults performed by different actors as equal.
After setting the requirements for the proof of concept, the research kicked off in November 2012 with a literature study. Based on the literature study a general design methodology for creating an automatic classification system was set up. One of the required demands for developing an automatic classification system is the need for a database of example recordings. These are used to train and test the classification system. Another step in the design of the system is to select a motion capture system to capture the vaults. During the literature study several motion capture techniques where studied, including inertial sensors and markerless video analysis. The study resulted in a preference for markerless video analysis over inertial sensors to serve as information source for the ACSV. Main reasons where that markerless video analysis is non-invasive to a gymnast, which means that unlike inertial sensors no sensor needs to be worn by the gymnast during the vault, and a database of video data was at hand during the start of the project. In addition to this, both ISL-DB and VU where confronted with an abundance of video data which where classified and stored manually in a database. An automatic classification system would be a major timesaving addition to the video system.
The markeless video analysis system used is known as the Gymnastics Coach CockPit developed among others by VU. The system is based on high speed video recordings and is implemented in the gymnastics hall of Flik-Flak Den Bosch (a top Dutch gymnastics association) and used during the World Championships in gymnastics of 2010 (WC 2010). The recordings of the WC 2010 comprise the database of training and testing samples used in designing the ACSV.
After adapting the scope of the project from inertial sensors to video analysis and receiving the WC 2010 database the development of the ACSV started in February 2013. To guarantee the proof of versatility the classification of a complete vault was segmented into a consecutive classification of four individual vault sections, namely:
- type of vault, or vaulting technique like Handspring or Tsukahara.
- number of somersaults performed in the second flight phase.
- type of somersaults, or posture during somersaulting.
- number of twists performed in the second flight phase.
Each vault section individually exists of several motion classes. For example, the number of somersaults vault section exists of the motion classes 0, 1 or 2 somersaults. The combination of the individual classification of the vault sections comprises the complete classification of the vault. Segmenting the vault classification into vault section classifications simplifies the classification problem by generating a lower amount of vault section motion classes compared to the original amount of vault classes.
By means of the video analysis characteristic vault properties, or so called features, are derived from a vault recording. These features comprise a unique pattern associated with the corresponding vault. To make the division between the different vault classes statistical algorithms known as a classifiers are used. The classifiers are trained by using the training samples from the database. The combination of the features and classifiers determines the sensitivity and robustness of the system. During the research several features and classifiers have been tested on their ability to classify the test samples. The percentage of test samples that are classified is known as the classification rate (measure of versatility). The percentage of correct classification of the test samples is known as the accuracy (measure of sensitivity as well as robustness). Per individual vault section the feature-classifier combination was chosen that scored highest in classification rate and accuracy.
Subsequently the complete classification of a vault was tested by combining the individual classifications of the vault sections. This resulted in a classification rate of 70% with an accuracy of 90%. This implies that 7 out of 10 recordings qualify for classification and 90% of these recordings are correctly classified.
Far out the most problems in classifying vaults occur in classifying the number of twists. Most gain can be achieved by improving the classification of this vault section. OSE is in consultation with ISL-DB about further development of the ACSV and implementation of the system in the gymnastics hall of Flik-Flak Den Bosch.