*3.4. Final Classification*

For action classification a standard leave-one-out cross-validation procedure is adopted. In each iteration, one sequence is left out and matched with the rest of sequences based on *AR* vectors. An *AR* which resulted to be the most similar (or less dissimilar) to the one under processing indicates its class. Indications from all iterations are verified with the original action labels and the percentage of correctly classified objects is taken (classification accuracy). For matching, three different measures are applied, namely Euclidean distance [36], correlation coefficient based on Pearson's correlation [37] and C1 correlation based on L1-norm [38].

**Figure 3.** Line graphs showing normalized area values for actions presented in Figure 2. The X axis corresponds to the consecutive frame numbers, while the Y axis corresponds to the normalized area values of the foreground object in each frame. Line graphs in the left column correspond to the actions performed in place (bending, jumping-jack, jumping in place, one-hand waving and two-hand waving), and line graphs for actions with changing location of a silhouette are depicted in the right column (jumping forward, running, galloping sideways, skipping and walking).
