Research On Dynamic Gesture Tracking Algorithm Based On Information Fusion Filter

As a natural interaction method,gesture interaction is an important basic research in the field of human-computer interaction.With the popularity of new somatosensory peripherals and the development of optimal estimation theory,it has become a key issue in the field of natural human-computer interaction that how to capture and track 3D gesture motion data quickly and accurately based on vision and understand the semantics of dynamic sign language in real time.In addition,the hearing-impaired people are at the top of the five major disabilities.Sign language is the most important way for deaf-mute people to communicate with the outside world,and the development of new,natural,friendly and portable gesture interaction system plays a vital role in teaching and communication of deaf people.Therefore,this paper focuses on the dynamic gesture detection and tracking of two key technologies in new human-computer interaction and dynamic sign language recognition.The human body part involved in the dynamic gesture movement is a typical non-rigid target.Due to the model is difficult to establish,it is difficult to accurately estimate and track the dynamic gesture movement trajectory.In the process of dynamic gesture tracking,the hand has arbitrary deformation,and the target gesture is not uniform motion in a certain direction,but there is a random movement with unknown speed in all directions,but there are random motions whose speeds are unknown in all directions,that is,dynamic gesture tracking has a distinctive feature:mobility.In addition,during the continuous movement of dynamic gesture,due to the number of possible uncertainties of the target and interference of environmental background and other factors,the system has certain randomness and uncertainty.Therefore,there may be four kinds of problems in the process of dynamic gesture tracking:robustness problem of arbitrary deformation gesture tracking,skin-like interference problem,error recovery problem when the target tracking is wrong,and occlusion recovery problem when multiple targets block each other.Aiming at the four kinds of problems that may exist in the process of dynamic gesture tracking,this paper proposes a dynamic gesture tracking algorithm based on information fusion filtering.Firstly,Kinect2.0 is used to obtain RGB color video stream and depth video stream of dynamic gesture motion,to construct dynamic gesture database,and to conduct information preprocessing for them respectively;Secondly,it is proposed that an automatic gesture detector training method based on region-convolution neural network(R-CNN)and migration learning algorithm,and to achieve automatic and accurate detection of target gestures in complex backgrounds;Then,in order to describe gesture state space more accurately,two maneuver models are added on the basis of traditional gesture state space with only non-maneuver models.By combining skin color information and depth threshold information in the process of dynamic gesture movement to improve the robustness of dynamic gesture tracking algorithm,on this basis,it is proposed that an information fusion Kalman filtering tracking algorithm;Finally,Complete the system setup and performance test of the dynamic gesture tracking algorithm.In order to verify the real-time and robustness of the dynamic gesture detection and tracking algorithm,four kinds of problems in the dynamic gesture tracking process are simulated.The experimental results show that the proposed algorithm can effectively solve the four types of problems that may occur in the dynamic gesture tracking process,and can achieve higher detection and tracking accuracy.