Multi Spatio-temporal Co-occurrence Measures for Human Action Classification

This paper deals with human action classification by utilizing spatio-temporal (ST) co-occurrences between labels of video-words that are stored within ST correlo-grams. Mutual information based clustering method is employed to reduce the size of the vocabulary created from local descriptors. Multiple characterizations for human actions in videos are extracted from the correl-ograms that are used for human action classification. These include a highly discriminative co-occurrence vector and a Haralick texture vector. The proposed method is implemented using a SVM classification technique. For evaluation purposes, the KTH and UCF Sports action recognition datasets, are used as they are the most well known and challenging datasets. The proposed method succeed in classifying different action classes, and improves the classification rate obtained by standard bag-of-video-words approach.