UBC Theses and Dissertations
Template-based action recognition : classifying hockey players’ movement Wu, Xiaojing
Although action recognition is remarkably easy for people, it is a difficult task for computers. A moving camera that takes broadcast-quality videos makes this even more difficult. This research focuses on actions in medium field. That is, a typical figure has a resolution of dozens of pixels in each dimension. The system is demonstrated using videos of ice hockey sport. An approach that breaks the problem into three sub-problems is taken. Figures of hockey players are first tracked with a self-initializing tracker. The stabilization process then refines the rough estimates about scale and position of a figure given by the tracker. Taking the stabilized results given by the stabilization process, the action recognition system uses motion and pose features to classify actions. A new stabilization algorithm is developed. The method uses a mixture of templates to estimate the position and scale of a figure. It helps to alleviate some of the accuracy and consistency problems. The consistency of the template library is addressed with a procedure that iteratively selects templates to better fit the training data. Our method is shown to consistently outperform a typical approach that uses only the best match with a set of synthetic image sequences. The research makes novel use of image gradients. It decomposes image gradients into four non-negative components. The decomposed image gradients (DIGs) are used to characterize poses. Quantitative performance comparisons are made between methods that use motion and pose features. The experiments show clear evidence that, for the kind of data that this research is interested in, pose features are better than motion features in terms of classification accuracy. The pose features are also superior to the motion features in terms of computational efficiency.
Item Citations and Data