Extending evolutionary Fuzzy Quantile Inference to classify partially occluded human motions

This work presents a framework that combines the concept of Fuzzy Quantile Inference (FQI) with Genetic Programming (GP) in order to accurately classify real natural 3d human Motion Capture data. FQI is a generalization of Fuzzy Gaussian Inference. It builds Fuzzy Membership Functions that map to hidden Probability Distributions underlying human motions, providing a suitable modelling paradigm for such noisy data. Genetic Programming (GP) is used to make a time dependent and context aware filter that improves the qualitative output of the classifier. Results show that FQI outperforms a GMM-based classifier when recognizing six different boxing stances simultaneously, and that the addition of the GP based filter improves the accuracy of the FQI classifier significantly. A mechanism allowing the FQI extended framework to deal with occluded data reasonably well is also integrated.