A fuzzy probabilistic inference methodology for constrained 3D human motion classification

Enormous uncertainties in unconstrained human motions lead to a fundamental challenge that many recognising algorithms have to face in practice: efficient and correct motion recognition is a demanding task, especially when human kinematic motions are subject to variations of execution in the spatial and temporal domains, heavily overlap with each other, and are occluded. Due to the lack of a good solution to these problems, many existing methods tend to be either effective but computationally intensive or efficient but vulnerable to misclassification. This thesis presents a novel inference engine for recognising occluded 3D human motion assisted by the recognition context. First, uncertainties are wrapped into a fuzzy membership function via a novel method called Fuzzy Quantile Generation which employs metrics derived from the probabilistic quantile function. Then, time-dependent and context-aware rules are produced via a genetic programming to smooth the qualitative outputs represented by fuzzy membership functions. Finally, occlusion in motion recognition is taken care of by introducing new procedures for feature selection and feature reconstruction. Experimental results demonstrate the effectiveness of the proposed framework on motion capture data from real boxers in terms of fuzzy membership generation, context-aware rule generation, and motion occlusion. Future work might involve the extension of Fuzzy Quantile Generation in order to automate the choice of a probability distribution, the enhancement of temporal pattern recognition with probabilistic paradigms, the optimisation of the occlusion module, and the adaptation of the present framework to different application domains.