Hand Gesture Recognition Based On Kinect Depth Super-pixels

With the rapid development of computer technology,communications between people and machines become more and more natural.In the field of human-computer interaction,hand gesture recognition plays an important part.The original hand gesture recognition employs data gloves or similar devices to collect gesture data.Because of its complex structure and high cost,it is difficult to for ordinary use.One the other side,vision-based methods have a high demand on environment.The depth camera(like Kinect)can eliminate the complex background factor.It could enhance the robustness of gesture recognition.Therefore,hand gesture recognition is becoming a research focus for scholars.In hand gesture recognition,a sharp increase in the number of templates will lead to a low recognition speed.The determination of the key templates is crucial.This is a difficult problem of clustering with unknown number of classes.We propose an approach-automatic key templates selection for hand gesture recognition with depth super-pixels.We use a weighted inter-intra similarity ratio to measure two factors: the compactness in one class,and the scatter between classes.An automatic hierarchical clustering algorithm is designed to select the key templates.Meanwhile,Earth Mover's Distance(EMD)is used to measure the similarity between inputs and templates.Experiments show it dramatically increases the recognition accuracy.The main work in this paper is summarized as following:First of all,we need to obtain information including colors,depths and joints through Kinect.According to the palm joint,required hand is located.We utilize the depth value of wrist joint as the threshold to segment the corresponding hand gesture.Secondly,the depth super-pixels is calculated through depth and color clustering.We implement EMD algorithm to compute SP-EMD distance between hand gestures.At last,hierarchical clustering algorithm is implemented together with above methods to select the key templates automatically.The experimental results show that the proposed methods can significantly reduce the number of templates by about 80%,and approximately improve the speed of system by 5 times,while keeps a similar recognition accuracy.