Action Categorization in Soccer Videos Using String Kernels

Action recognition is a crucial task to provide high-level semantic description of the video content, particularly in the case of sports videos. The bag-of-words (BoW) approach has proven to be successful for the categorization of objects and scenes in images, but it´s unable to model temporal information between consecutive frames for video event recognition. In this paper, we present an approach to model actions as a sequence of histograms (one for each frame) represented using a traditional bag-of-words model. Actions are so described by a string (phrase) of variable size, depending on the clip´s length, where each frame´s representation is considered as a character. To compare these strings we use Needlemann-Wunsch distance, a metrics defined in the information theory, that deal with strings of different length. Finally, SVMs with a string kernel that includes this distance are used to perform classification. Experimental results demonstrate the validity of the proposed approach and they show that it outperforms baseline kNN classifiers.