Research On The Key Technology Of Vision-based Static Hand Gesture Recognition

With the development of information technology,as a great invention,computers are impacting people's daily life deeply.As an important application of computer technology,the natural human-computer interaction technology based on biometric identification has a close relationship with people.The biometric identification is a technology which uses the computer technology to process images or videos,trying to extract the specific biological features and recognize the living body.The technology is becoming a hot research field of artificial intelligence.Human-computer interaction using the biometric identification technology has some advantages,such as convenience,uniqueness etc.Some common used biological features include faces,fingerprint,iris,hands and so on.Hand gesture recognition as a kind of human-computer interaction technology attracts more and more researchers' attention by its nature,convenience,abundant information.But hand is uncertain and multiple,so there are still some challenges to solve.Hand gesture recognition is becoming a hot and hard research field of human-computer interaction technology.A gesture recognition system mainly includes three parts:preprocessing,feature extraction and classification.This thesis studies on some algorithms of static hand gesture recognition,and it focuses on feature extraction and classification.This thesis has completed the following tasks:Firstly,the thesis introduces some classical algorithms of feature extraction and classification.The theories,steps,weaknesses and strengths of them are summarized.Secondly,for the low recognition rate and high feature dimension of the basic Local Binary Patterns(LBP),this thesis proposes a multi-neighborhood weighted fusion local binary patterns method.It is an improved algorithom of the basic LBP.It calculates two code images by processing the two layers of neighbourhood using different ways respectively.And it gets the 256-dimensional histograms of the two code images.Then it uniformly quantizes the two 256-dimensional histograms to 32-dimensional histograms.At last,it makes weighed fusion of the two 32-dimensional histograms and the final 32-dimensional histogram is the feature vector.Experimental results which are done on two gesture databases show that the proposed method can increase the hand gesture recognition rate and reduce the feature dimension drastically.Thus it can improve operating speed.Thirdly,this thesis researches the Non-Negative Matrix Factorization(NMF)and the Compressive Sensing(CS).We use them to design a hand gesture recognition system.Original images are projected to low-dimensional subspace by using NMF,then gestures can be recognized by the classifier which is designed by CS theory.The experimental results which are done on two gesture databases show that the CS classifier performs better than some other classifiers no matter on the recognition rate or resisting occlusion.The NMF also has a better occlusion resistance than the Principal Components Analysis(PCA).