Kinect Gesture Recognition Method And Application Through Decision Tree Classification And Trajectory Fitting

With the development of computer technology,human-computer interaction(H-CI)model has undergone tremendous changes.The traditional mouse and the graphical interface have been unable to meet the growing demand for interactive models,there-fore more natural interaction patterns emerge.Now,users can interact with computer via the approaches such as voice,touch,gesture etc,which are probably more ac-cordant with human daily habits.In these techniques,gesture interaction technology is widely applied in home entertainment systems.Moreover,as an innovative application,Kinect has provided an important equipment support for gesture interaction researches in the past several years.However,there still exist some shortcomings and problems in vision-based hand gesture recognition.For example,the case of hand tilt will affect the current methods,typically consisting of template matching algorithms.When there exists a large number of templates need to match,the time of matching will probably be increased greatly.Furthermore,the complex background and scene environment are always susceptible to hand gesture recognition.Against these problems,this paper proposes a Kinect gesture recognition method which combines decision tree and trajectory fitting to explore the static gesture recog-nition and dynamic trajectory recognition problems.Firstly,we introduce the structure and the key technologies of Kinect,and implements the extraction of palm contour by using skeletal tracking from the the depth images generated by Kinect.Secondly,a cir-cumference sequence curve is constructed for each palm contour,and extreme points are used to cut out the contour of each finger.Thirdly,we design the feature sets of gesture contours,and train decision-tree based classifiers on these feature sets,in or-der to achieve the purpose of the static gesture recognition.We have evaluated this method on two datasets consisting of our own dataset and the 10-Gesture dataset.Our experimental results show 85.31%and 85.30%mean accuracies on these two datasets,respectively.To recognize dynamic gesture,we further combine straight-line fitting and parabo-la fitting to assist recognizing gesture.This method recognizes gestures by using hand trajectory to avoid hand pose noises.We evaluated this method on our own dataset,averagely achieving 97%for straight-line fitting and 92%for parabola fitting.Static gesture recognition and dynamic gesture recognition are thus combined together to un-derstand the type of a moving hand pose with the average computational time of 0.251 second.We apply this hand-based interaction recognition method in our FreeScup plat-form,which aims at assisting sculpture design,and achieve real-time performance on interacting with 3D sculpture models.