Dynamic Gesture Recognition Based On Kinect

Gesture has the advantages of convenience, nature and intuition. It has been increasingly used in the field of human-computer interaction. Compared to static gesture, dynamic gesture is rich and flexible. Therefore, this paper mainly focuses on dynamic gesture recognition.Dynamic gesture recognition mainly includes the following steps: obtaining palm position sequences, encoding trajectory features, learning the model parameters and thresholds, classifying gestures. Human skeleton joints’coordinates can be obtained in real time through the Kinect which is an equipment released by Microsoft. The next step is to encode gesture features using K-means. But this method is difficult to distinguish undefined gestures with predefined gestures. This paper presents a spherical direction discretization (SDD) method to encode gesture features and features encoded by this method can show the differences of each gesture. Then this paper introduces the dynamic time warping (DTW), the longest common subsequence (LCSS) and the hidden Markov model (HMM) in dynamic gesture recognition. Combining KNN with LCSS solves the different habits of users. A statistical method is used to calculate the threshold, which is applied to classify the predefined gestures, automatically detect and filter out undefined gestures in HMM-based dynamic gesture recognition method。The experiment results show that SDD+KNN+LCSS outperformed K-means +LCSS by 13.45% F1-score. SDD+ DTW outperformed K-means+ DTW by 3.08% F1-score.SDD+ HMM can get 81.54% F1-score。...