Spatio-Temporal Features

The features we use are based on the simple Haar basis features used by Viola and Jones[3]. We have extended them into the temporal domain as shown in figure 2. We rely on the learning system to combine these basic features in such a way as to depict the motion rather than choosing complex classifiers to start with. Each feature returns a value which is the difference between the values of the two volumes. Since all of our features are rectangular volumes we can speed up the calculations if we consider an interior and exterior box. The value of the feature is then given by the summation of the pixels in the blue box minus twice that of those in the red box.

figure 2 - Five of the seven features used are shown, the two not shown are the inverses of (a) and (b)