A space-time SURF descriptor and its application to action recognition with video words

A novel method to human action recognition is presented with the combining of a new space-time Speeded Up Robust Features (SURF) descriptor and the bag of video words (BOVW) approach. In our method, we have extended the SURF so that it can better represent the inherent spatio-temporal information of the video data for action recognition. To utilize this descriptor in the action recognition framework, the BOVW schema with a soft-weighting strategy is exploited. Experiments, conducted with the KTH´s action recognition dataset, have shown that the proposed method can achieve an outstanding performance in both computing speed and accuracy contrast to the traditional methods.