Research And Implementation Of Dynamic Gesture Trajectory Recognition Technology

With the development of virtual reality,human-computer interaction has undergone a continuous process from command line,graphical interface,to natural user interface(NUI).NUI refers to as a method in which users can interact with computer in the most familiar and natural way,such as gestures,voice,touch,and so on.Served as a significant form of human-computer interaction,the hand gesture,in particular,can release users from the restraint of traditional mouse and keyboard,which is more acceptable to humans.Therefore,wide application prospects could be seen on NUI.The existing gesture recognition technology can be divided into static gesture recognition and dynamic gesture recognition.Recognition of dynamic gesture can be regarded as a continuous trajectory of time sequences in the model space composed of a series of static gestures.Being as a subclass of dynamic gesture recognition,trajectory of gesture recognition emphasizes real-time from the dimension of time,while it mainly tracks the gestures trajectory.The real-time human-computer interaction system is needed for most applications of dynamic gestures.Therefore,with a high research value,trajectory of gesture recognition of real-time has the widest range for application.However,from the current studies in this area,well developed theories and models are relatively inadequate,while the existing methods are deficient in recognition accuracy and with low operating efficiency.Hence,in terms of the problems existing in this field,two recognition algorithms that applied to two different scenes are proposed in this thesis,which achieve benign results during the process of recognition.The major innovations of this thesis can be summarized as follows:First of all,an Adaptive Dynamic Time Warping algorithm is proposed by establishing an adaptive state space model hence the gesture recognition can be well adapted to the possible errors in the process of gesture recognition and a parameter optimization scheme is designed to effectively improve the performance of the algorithm by adding global constrain conditions,slope matrix and elimination threshold optimization.The experiment results show that this method is better than the traditional DTW and slightly lower than HMM in recognition aspect.In efficiency aspect our method is much better than HMM and equals DTW.Then,we find that ADTW's method is not particularly ideal in terms of accuracy and time,for gradually expanding gesture template library.From the user's view,it can give users feedback and prompting in the process of recognition,which is a good direction for optimization.Therefore,this research is to use the framework of particle filter for state model,and the author redesigns the feature variables for particle filtering process,a resampling algorithm proposed by improving the importance sampling process.According to the feedback of experimental results,this method is obviously better than the existing algorithm both in recognition accuracy and efficiency.Finally,based on the two methods,we proposed in this paper,multiplatform gesture trajectory recognition system is designed to intuitively verify the feasibility and rationality.The system includes a whole process from gesture collection to recognition,compatible with multi-platform,has a certain practical value.