Kinect-based Prediction Technology For Falling Risk Of Virtual Reality Users

In order to reduce the risk of falls caused by VR dizzy in immersive virtual reality(VR)users,most of the current researches use sensors to collect acceleration characteristics and predict fall through model matching and threshold method.However,in practical applications,the method based on the acceleration characteristics often has the problem that the sensor is worn tediously,which affects the user experience and the lack of posture information leads to a low recognition rate.In addition,due to a single feature dimension,current fall prediction methods often need to sacrifice a certain amount of prediction time to improve the recognition accuracy,resulting in insufficient margin time.This thesis proposes a fall prediction method for virtual reality users,aiming at the problems of low recognition rate,insufficient margin time and unsuitable for immersive VR scenarios in the existing fall prediction methods.The main research content is as follows.Firstly,for the current device based on the acceleration feature,the fall prediction method is cumbersome to wear and it is not suitable for the immersive VR scene and the problem of insufficient margin time.A method of HMM-DT fall prediction based on joint angle characteristics is proposed.Based on Kinect skeleton joint data,this method first uses an improved joint angle feature to simplify the data dimension and fully describe the human posture;then it performs predictive research on the human fall process and improves the existing fall prediction based on HMM sequence matching.The method,combined with decision tree(DT)classification theory,takes each feature as a branch of the decision tree,and finally results in a classification prediction.The experimental results show that this method achieves the goal of predicting falls in immersive VR scenes,significantly prolonging the margin time,and reducing the risk of falling subjects.Secondly,aiming at the existing overfitting phenomenon in the HMM-DT fall prediction method based on the joint angle feature,resulting in a high misjudgment rate,a multi-dimensional feature-based HMM-RF fall prediction method is proposed.The method improves the existing method from two aspects.Firstly,the angular velocity feature is added to the joint angle feature to represent the timing state.Secondly,the idea of constructing the decision tree in the original method is improved to random structure by adopting the idea of random forest(RF).In forests,the characteristics of “multiple votes for winners” of random forests are used to eliminate the influence of overfitting in the original method.Experiments show that this method has a significant improvement in the recognition accuracy and recall rate.Third,based on the above research,we designed and completed a fall prediction system in a virtual reality scenario to analyze the user's gestures in real time to predict whether they would fall;and fed back the prediction results to the system interface and virtual reality scene.