Appearance-based hand gesture identification

From the moment of human beings coming to earth, hands are the most dexterouspart on our body. Our ancestors use their hands to support their lives and manufacturetools. Also, human use hand gestures to communicate with each other as asupplementary of our speaking language. With the advancement of computer visiontechnologies, the hand functionalities do not restrict on the above areas. Instead ofusing buttons, joysticks, keyboards, mice and so forth to interact with the intelligentdevices, scientists attempt to implement these functionalities solely by our hands.This thesis is concerned with studies and developments of generating an effectiveappearance based gesture recognition system that can be adapted to different complexbackgrounds. The problems raised in this research are challenging due to the highlyvariable hand skin color, geometric features and complex backgrounds.The problems are tackled by beginning with an iterative training scheme for segmentinghand area from input image sequences. This segmentation method is developedfrom integrating several commonly used methods to synthesize the merits ofthese methods. Then, a self-organizing neural network is employed for recognizinghuman hand gestures. More specifically, with the application of the growing neuralgas, a limited number of neural nodes is allocated over the segmented hand sub-image.By analyzing the distribution and connection of these neurons, the semantic meaning of gesture can be obtained.The effectiveness of the proposed method is verified by simulations and experiments.It is shown that the proposed methods are suitable for the appearance basedhuman gesture identification. To start with, several hand images which are obtainedin different scenarios are used in the test, and results show that the method can segmenthand areas with a satisfactory accuracy. Then, with regard to the proposedself-organizing neural network for gesture recognition, experiments are further conductedto prove that high computational efficiency and low error quantization can beobtained.v