Hand gesture recognition methods based on concept learning

Abstract

Human Computer Interaction (HCI) plays an important role in the digital age. HCI applications allow people to interact with digital devices such as computers in different HCI modalities such as keyboard, mouse, touchpad and hand gestures. Hand gestures are an attractive HCI modality in some applications such as gaming and assisted living, but recognising hand gestures is not easy and has been the subject of research for many years. The key problem in hand gesture recognition is how to learn to recognise different hand gestures under varying conditions in that the same hand gestures may be performed in different shape, duration and lighting and consequently the sensor data is usually very noisy. Therefore, hand gesture recognition is a considerable challenge for the research community.

The overall aim of our research is a concept learning solution for hand gesture recognition. This thesis proposes a concept learning framework with three modules – concept grounding, concept association and concept adaption. Several new methods are proposed under the framework. The first one is a data mining fusion method CEMF, i.e., Clustering Ensembles Based on Multi-classifier Fusion, which uses the idea of multi-classifier combination to solve the clustering consensus problem and to provide as the basis of the concept learning framework for hand gesture recognition.

The second method is a hybrid approach to hand gesture recognition. Itgenerates frequent patterns from the hand gesture training data set by usingassociation rules mining in order to recognize some incomplete hand gesturesequences.

The third method is an adaptive algorithm that employs the idea of immune algorithms to build the adaptive model to adapt the gesture concepts to different users to recognize the gestures of different users. Experimental results show the effectiveness of the concept learning framework presentedin this thesis.

Date of Award	May 2015
Original language	English
Supervisor	Jun Liu (Supervisor) & Hui Wang (Supervisor)