| Summary: | ela uba a m trac in ing ew the erim anc a great deal of attention in re-umerou terface n comp action f the i otions ecent y used a, refer proach In particular, the bag-of-features framework (Csurka et al., 2004) has been successfully applied to motion recognition (Dollar et al., 2005; Laptev et al., 2008; Wong and Cipolla, 2007) as well as image recognition (Bosch et al., 2007). In that framework, the rec-ognition of motion relies on local features which are based on sim-ple histograms of spatial gradient orientations (HOG) (Dalal and Triggs, 2005) and space–time derivatives (Dollar et al., 2005; Zel-ture the space–time shape of the motions and do not have much discriminative power. Therefore, in the bag-of-features framework, XYT domain, by developing the gradient local auto-correlation for image recognition (Kobayashi and Otsu, 2008) to extract space– time motion features. The local relationships correspond to geo-metric characteristics, i.e., gradients and curvatures, which are fun-damental properties of space–time motion shape. For motion recognition, we utilize the frame-based features which are ex-tracted from sub-sequences sampled at dense (grid) time points along the time axis. In this approach, referred to as the bag-of-frame-features approach, the frame-based features sufficiently characterize the motion in the spatial domain in contrast to the lo-cal features, and the motion in the entire sequence is described by
|
|---|