Study On Dynamic Gesture Recognition Algorithm Based On Dictionary Learning

Dynamic gesture recognition technology based on vision uses the gesture as the input of the computer directly,which gets rid of the shackles of the traditional contact-type devices and achieves a friendly way of human-computer interaction,so it has been widely concerned by domestic and foreign researchers.However,due to the fact that the technology involves many subject areas,and the gesture itself also has the characteristics of diversity,complexity,and difference in time and space,so it still faces enormous challenges in practical applications such as somatosensory games.Aiming at how to improve the accuracy and real-time performance of gesture recognition,gesture segmentation,gesture tracking and gesture recognition are studied respectively.In the stage of gesture segmentation,the gesture is segmented using the combination of skin color information and motion information,which effectively solves the interference of faces,arms and skin-like regions.In the stage of gesture tracking,first,aiming at the limitation of Camshift(Continuously Adaptive Meanshift)cannot actively search target,and is difficult to adapt to target tracking at high speed,the method is combined with the background difference and the Kalman filter algorithm to track the gesture,which can solve the problem of losing the tracking goal when gesture moves too fast or there is a large area of skin color interference and gesture disappears;Then,use two-time static method to determine the starting and ending position of the gesture motion,and separate the predefined meaningful gesture and meaningless gesture.In the stage of gesture recognition,this thesis applies LC-KSVD(Label Consist K-SVD)dictionary learning algorithm to dynamic gesture recognition based on vision for the first time and achieves a dynamic gesture recognition method based on dictionary learning: aiming at the problem of different gesture length,each dynamic gesture sample is represented as a column vector and does normalized linear interpolation processing,and the recognition problem of gesture is converted to sparse representation problem on the overcomplete dictionary,which reduces the computing cost and recognition time,satisfies the real-time requirement;Aiming at how to increase the distinction of different gesture classes and reduce the difference of the same gesture classes,this thesis adds a class label for each gesture sample and closely associates the label information with the atomic items of the dictionary through a predefined discriminative sparse coding so that gestures of the same class have similar sparse codes and those in different classes have dissimilar sparse codes,which strengthens the accuracy of gesture recognition.To verify the effectiveness of method in this thesis,test on self-defined 10 types of dynamic gesture set.The experiment is divided into two stages: training and testing.In the training stage,a compact and optimized overcomplete dictionary is obtained by the LC-KSVD algorithm;In the testing stage,solve sparse codes of the gesture vectors to be tested on the overcomplete dictionary,then send sparse codes into classifier to recognize.Experimental results show that compared with some existing recognition methods,recognition method in this thesis can greatly reduce the recognition time while ensuring the recognition accuracy.