Multi-prototype fuzzy clustering with fuzzy K-nearest neighbor for off-line human action recognition

Fall detection of elderly in home environment is an important research area. The fall detection is a part of the human action recognition. In this paper, a human action detection using the fuzzy clustering algorithm with the fuzzy K-nearest neighbor from view-invariant human motion analysis is implemented. In particular, the Hu moment invariant features are computed. Then principal component analysis is utilized to select the principal components. The fuzzy clustering algorithm (either fuzzy C-means, Gustafson and Kessel, or Gath and Geva) is implemented on each class to select the prototypes representing the class. From the results, we found that the best classification rate on the validation set is around 99.33% to 100%, and the classification rate on the blind test data set is around 90%. We also compare the result from fuzzy K-nearest neighbor with that from K-nearest neighbor. The fuzzy K-nearest neighbor result is better as expected.