Prototype features

The feature space of $m \times m$ motion histograms is still too large. To detect potentially important feature signals we apply vector quantization to the histogram feature vectors classifying them into a dictionary of $K$ prototype features, $\mathbf{P} =\{p_1, \dots, p_K\}$ using K-means (figure 3).

Figure 3: Create prototypes from set of features

Note that using this discrete quantization, it is possible that two image frames with similar features will have different prototype labels. In this case, the two prototypes must also be similar. This problem is resolved, as we will see later, by keeping track of similarity among the prototypes.