Bag-of-words with aggregated temporal pair-wise word co-occurrence for human action recognition

The bag-of-words (BoW) representation has successfully been used for human action recognition from videos. However, one limitation of the standard BoW is that it ignores spatial and temporal relationships between the visual words. Although several approaches have been proposed to deal with this issue, we propose an extension which is arguably simpler yet quite effective. The proposed representation, t-BoW, captures only temporal relationships between pairs of words in an aggregated way by counting co-occurrences at several temporal differences. Unlike other approaches, neither spatial nor hierarchical information is accounted for explicitly, and no significant change is required in the quantization or classification procedures. Performance improvements over the traditional BoW and other BoW extensions are experimentally observed in the KTH, the ADL, the Keck, and the HMDB51 action/gestures datasets