Classifying 3-dimensional point light actors' gender using structure-adaptive self-organizing map

As a basic step to understand human behavior, we search for an optimal classifier to recognize the biological movements using moving point lights attached on actor's bodies. Classifying the patterns with self-organizing map often fails to get successful results with its original unsupervised learning algorithm. This paper presents a structure-adaptive self-organizing map (SASOM) which adaptively updates the weights, structure and size of the map, resulting in remarkable improvement of pattern classification performance. Two physical input features of the movement patterns have been used: positions and velocities of six locations. We have compared the results with those of conventional pattern classifiers and human subjects by obtaining the recognition accuracy, discrim-inability and efficiency. SASOM turns out to be the best classifier producing 97.1% of recognition rate on the 312 test data from 26 subjects.