Toward A Biologically-Inspired Representation of Human Affect

We propose a method to reveal the features used by humans in the classification of human movement and apply it to the case of classifying arm movements as angry or happy. The method begins with psychophysical experiments investigating the human classification of point-light movements. Then by comparing the results from these perception experiments with the results of principal component decomposition, we can find a particular feature component that has the highest correlation with human perception. In order to verify the component, we reconstruct movements by using either the first two PCA components alone as well as those two components and the feature that correlates highly with human perception. Finally, we used a Parzen window to test the recognition effectiveness of these reconstructed movements