Research On Hand Gesture Recognition Based On Machine Learning

With the continuous development of computer graphics and virtual reality technology,the requirements for human-computer interaction are getting higher and higher.Traditional gesture recognition is based on camera photos and 2D gestures for recognition and classification.The extracted features is often difficult to accurately represent the difference between different static or dynamic gestures.For the problems of high degree of freedom and inaccurate feature representation of different gestures in static gesture recognition,complex timing sequence,spatial variability and inaccurate feature representation of dynamic gestures,etc.In recent years,feature extraction,feature fusion,feature dimensionality reduction,support vector machine and other methods applied to gesture recognition have become a research hotspot.It is difficult to extract the complete gesture's features from the gesture's data and classify them effectively.The gesture recognition method based on gesture geometry features and integrating multiple features is one of the powerful tools to solve this problem.The main work of this paper and the research results obtained are as follows:1)In order to solve the issues of high degree of freedom and inaccurate feature representation of different gestures in gesture recognition,a new hand gesture recognition method is proposed here,which combines the joint rotation feature and fingertip distance feature.Firstly,the 3D position information of 20 hand joints is calculated from the depth map of the underlying hand by using hand template.Then the quaternion joint rotation feature and the fingertip distance feature are extracted by using the position information of the hand joints,which constitutes the intrinsic representation of the hand gesture feature.Finally,the hand gesture can be effectively recognized and classified by using support vector machine(SVM).Experiments show that this feature representation can effectively on behave of different gestures,and it can achieve a higher classification and recognition accuracy in static gesture recognition.2)In order to solve the problems of complex timing series,spatial variability and the inaccurate feature representation of different dynamic gestures,a novel dynamic hand gesture recognition algorithm is proposed here by combining the global gesture motion and local palm motion.Firstly,based on the given hand joint positions,some data preprocessing steps are performed for dynamic gesture data,such as removing of the invalid gestures frame,completing the gesture dataset,and normalization of the joint lengths for different gestures.Secondly,the key frames will be extracted according to the distance function defined by the difference of hand translation and rotation,fused by the difference of panning and rotating of palm fingers.Thus,according to the extracted key frames,the global gesture motion can be defined as the distance between the hand principal direction vectors and the displacement of the hand center point,whilst the local palm motion can be calculated as the angle difference of joint rotation quaternions and the relative distance of fingertips.Finally,by combining the global gesture motion and the local palm motion,the dynamic gestures can be classified by using the LDA(linear discriminant analysis)and the Gaussian kernel based SVM(support vector machine).Experimental results show that this method can effectively extract key frames from dynamic gestures,and replace all dynamic gesture frames with key frames,thus avoiding data redundancy.Based on the characteristics of static gesture and dynamic gesture,this paper effectively combines the geometric characteristics of gesture and the continuity of time and space to achieve effective gesture recognition.In this paper,on the basis of describing the research status of gesture recognition at home and abroad,the specific theoretical model and corresponding gesture recognition algorithm are proposed,and the experimental results of the algorithm and the comparison with other existing methods are given.