Research On Hand Detection And Tracking

With the rapid development of computer technology and the popularization ofpersonal computers, and the arise of The Internet of Things, Speech Recognition,Artificial Intelligence, Virtual Reality technology, Human-Computer Interactionbegan to popular, which mainly through object recognition and object tracking tointeract with computer. Human-Computer Interaction allows users using voice, facialexpressions, gestures and more natural way of human communication habits forcomputer operations, so this field had been greatly researched in recent years.Since the gesture is most used in HCI(Human-Computer Interaction) and also ismost convenient way for human to interact, the Gesture Interaction become the mostimportant Interaction in Human-Computer Interaction. Hand detection and Handtracking is the first step to achieve Gesture Interaction, because we can only analyzethe meaning of hand after we detect it and track it. More and more scholars focus onresearching on the detection and tracking of the hand.Hand detection is a kind of object detection, and its main purpose is to detectthe hand in the video sequence. Common hand detection methods are Hand detectionby Motion Information, Hand detection by Image Segmentation, Hand detection byTemplate matching and Hand detection by Statistical learning model. After analyzingthe advantages and disadvantages of these methods, this context use statisticallearning model to detect the hand. We use the improved Local Binary Pattern todescribe the hand region. First we use support vector machine to learn the hand LBPfeature of hand samples, after a period of learning, we will get a classifier of hand.Then we can use this classifier to detect the hand appeared in the video sequence.The experiment results shows that this method has a higher robustness and real-timetoo.After the completion of hand detection, we often need to track hand so as to getthe trajectory of the hand in the entire process of HCI. After analyzing the originOn-line AdaBoost algorithm, this paper propose a improved on-line AdaBoost algorithm to track the hand, this algorithm reduced the "tracking drift" phenomenonin origin algorithm. And the main method is: We using On-line AdaBoost algorithmand LK optical flow method based on consistency feature point to track the hand,then update the weak classifiers of On-line AdaBoost algorithm when the overlap oftwo algorithm hand region larger than specified threshold(70%is specified by thispaper), instead of updating the weak classifiers always. This method can reduce thebad hand samples produced during tracking, so it can improve the accuracy of handtracking.Meanwhile, in order to analyze the hand move action in monocular camera, webuilder a shape structure model(SSM) of hand, using this model we can analyze thehand operation and action in the entire video sequence, including hand translation,rotation and stretching. The experimental results show that this model has a goodrobustness and accuracy.