IIIT Hyderabad Publications |
|||||||||
|
Ontology and Event Grammar Based Analysis of TennisAuthor: Pushyami Rachapudi Date: 2017-12-26 Report no: IIIT/TH/2017/89 Advisor:Navjyoti Singh AbstractIn the thesis, an ontology and event grammar of tennis is proposed in order to enhance the methods of analysis in tennis. This work specifically focuses on the analysis required for improving the game from the coach’s point of view. Our ontological study has a vocabulary involving about 760 reusable terms related to tennis which are broadly divided into 12 top classes. The scope of this thesis deals primarily with three classes ‘play’, ‘score’ and ‘court’ and related terms out of the twelve. These terms and their relationships are defined with the help of event trees, score graphs and labelled images which are based on the rules legislated by International Tennis Federation. Any game of tennis involves a particular sequence of events. With the help of event trees, a generative event grammar of any play has been designed. This grammar generates all possible legal sequences of events that can take place in a game of tennis. However, a particular game is a unique sequence of such possible events. A game of tennis is represented as a finite state automaton, at any given point it can be in exactly one defined state. Annotated data is crucial in any form of analysis. We have created a real time annotator for tennis on the android platform based on the proposed event grammar. This integration of grammar into the application reduces the error rate and also makes the process efficient. The annotated file is then processed for generating play statistics at two levels. First, basic features such as success rates of shots/serves, number of forehands/backhands hit/miss etc. are analyzed. These features are very similar to those presented during the broadcast tennis videos in major tournaments although our annotation has more data like sub-events. Secondly, we infer sub-events between shots. Such detailed annotation points towards deeper inquiry as to why a shot was not executed properly, if it failed. An ergonomic and easy to replicate video documentation eight camera standard for the game of tennis is proposed. For deeper analysis, we create a data set of (singles) tennis games in accordance with the reduced version of the standard. Computationally two main features were extracted from this video data set: A. Player detection by a generalized feature based motion tracking algorithm, and B. Ball tracking by an improvised object-and-trajectory based algorithm. Every shot played is defined as a 18-feature tuple in the ontology. Selected features from the definition were computed by extracting ball trajectory and player position in real world situations. The features thus extracted are then classified and validated using SVM and KNN techniques. The features extracted along with their accuracies are: spin of the ball (83.85%), trajectory height (95.56%), position (94.65%) and number of bounces (98.93%). Additional application of shot prediction (88.75%), which predicts the next likely shot to be hit based on the player and ball bounce positions, was also implemented. An adaptive probabilistic model was developed to analyze a player’s game by predicting the conditional probabilities of an event based on a previously occurred event. A graph based on event trees is constructed to store the occurrences of a particular event and features. The data changes dynamically as the game proceeds and at any given point will give an experimental probability of the next likely events based on the current and past events. Proposed analysis reduces search space for coach to identify areas of improper execution and work towards improving player’s performance. Full thesis: pdf Centre for Exact Humanities |
||||||||
Copyright © 2009 - IIIT Hyderabad. All Rights Reserved. |