Research On Free Hand Tracking Based On Improved VLMM And Gesture Analysis

With the development of human-computer interaction, the interaction with natural,convenient, and harmonious are becoming the direction of development of human-computerinteraction technology. As a natural and intuitive way, human hand is playing a more andmore important role. For the deepening of the theory and application of intelligenthuman-computer interaction, especially in the aspect of virtual reality, vision-based free handtracking, has an important research value and application prospect.The essence of the model-based tracking method is to establish the correspondencebetween the characteristics of three-dimensional gesture model and the characteristics oftwo-dimensional frame image, and match the features of frame image obtained and handmodel in order to identify the best gesture model. However, a hand is a multi joint, complexnon-rigid structure. It is not only characteristic of high degree of free, but also characterizedby ambiguity, diversity, and differences in time and space, which makes the model-basedtracking method, on itself, is a very difficult thing. Besides, it is a challenging issue. Thisproject mainly relies on the research the real-time and robustness of particle filters for handtracking by the National Natural Science Foundation of China (No.60973093), and thepurpose is to research on three dimensional hand tracking method.In this paper, we make use of Probabilistic Graphical Models which is one of the tools ofthe sequence of semantic modeling of continuous dynamic characteristics, and think from thenew point of Newton’s First Law to research on hand tracking algorithm. The maincontributions include the following:(1) Production model for specific interactive tasks---variable length markov modelWith the establishment of a virtual assembly system, we obtain the gesture data fordifferent operators to complete the same interactive tasks, and then we see it as a training datato establish the variable length markov model. In the process of training, we choose gesture asa study object which means continuous posture in time order. In this way, it can both reducethe huge amount of computation due to the high degree of free, and reduce the probability ofthe deformity gesture to fully take into account the correlation between the joints.(2) A new algorithm based on variable length markov model for3D hand tracking is presented.A new framework combining particle filter with variable length markov model isproposed to solve the movement tracking of articulated hand model. By the variable lengthmarkov model, we can propagate particles only in plausible directions to reduce the numberof required particles. In this manner, we can reduce the time cost, and improve the trackingspeed. Simultaneously, we select PERM (Pruned-Enriched Rosenbluth Method) whichreplaced the standard resampling method in the resampling process. The PERM can avoid theloss of particles with small weighs, and also can inhibit the power to weight of particles arerepeated copied to enhance the diversity of particles. The experiment shows that our proposedalgorithm, compared with particle filter, can significantly improve tracking speed, betterensure the diversity of resampling particles, and improve the accuracy of the trackingalgorithm successfully.(3) A novel model according to Newton’s First Law is presented, named Inertial timemodel.Inspired by Newton’s First Law, we think hand movement is inevitable inertiacharacteristics in the process of interacting with the virtual assembly system. Virtual assemblysystem as the basic background, reducing the number of particles involved in randomsampling as the purpose, we analyze the movement characteristics for operators.Consequently, we discover the features of the stationary phase, and we propose and name themodel for the inertia time. In the running time of this model, there is no need to learn andsample particles, just moving accordance with the inertia feature. The longer the running timeof the model, the fewer the number of particles acquired and the fewer the movement learned,which contribute to reduce the time cost of the hand tracking algorithm.(4) A new algorithm based on inertial time model is put forward.In our algorithm, we take advantage of the inertial model to improve the algorithm basedon variable length markov model. Each state in the variable length markov model only needsto study typical gesture of beginning and end of the dynamic sequence, the other gestures canbe obtained from the inertial time model. This manner can reduce the number of particleswhen training the model, and better predict the behaviors outside the learning model. Inaddition, we introduce detection mechanism with reference to the sampling mechanism of the particle filter, in order to ensure the accuracy of the tracking. The experimental results showthat our algorithm can successfully predict the free movements of hand to improve thetracking accuracy of the non-learning, greatly reduce the number of samples to save the timecost, and narrow the range of forecast gestures samples in the process of tracking whichsettles the high dimension space searching problem.