In interest point based human action recognition, local descriptors are used to represent information in the neighbourhood around each extracted space-time interest point. The performance of the action recognition systems highly depends on the invariance and distinctiveness of the local spatiotemporal descriptor adopted. In this paper, we propose a new descriptor based on the Wavelet Transform taking advantage of its capability in compacting and discriminating data. We evaluate this descriptor on the extensively studied KTH action dataset, using the Bag-of Features framework. Results show the Wavelet Transform based descriptor achieves the recognition rate of 93.89%, which is better than most of the state-of-the-art methods.
|Published - Sept 2010
|BMVC 2010 - 21st British Machine Vicion Conference - Aberystwyth, UK
Duration: 1 Sept 2010 → …
|BMVC 2010 - 21st British Machine Vicion Conference
|1/09/10 → …