Research On Hand Gesture Recognition Based On RGB-D Images

As one of the important topics in the field of computer vision,hand gesture recognition has received extensive attention from researchers in various industries in the past few years.With the development of artificial intelligence technology,human-computer interaction(HCI)gradually penetrates into all aspects of industrial production and people's daily life,while research on hand gesture recognition,which is one of the most important means of HCI,is of more practical significance.The emergence of Kinect sensors provides new ideas for the research of hand gesture recognition.In this paper,based on the RGB-D images collected by Kinect sensor,the method of hand gesture recognition is studied.From three aspects,the feature extraction methods of hand gesture are innovated,and the matching method is improved to increase the accuracy and efficiency of recognition.In this paper,the process of preprocessing the images containing hand gestures is first employed.The depth information in the depth images collected by the Kinect sensor is used to segment the hand gesture region and remove the influences of cluttered background and noise,and then only the hand gesture area is preserved.The Canny operator is used to extract the contour of gestures.The segmented images with clear gesture contours provide convenience for the following process of gesture feature extraction.Then,a finger-emphasized multi-scale descriptor is proposed.According to the particularity of hand gesture contours,a finger-emphasized factor is proposed to emphasize the contour points of the finger part and highlight the feature information.Then the multiscale description of hand gesture contour is constructed.In the experiment period,recognition based on dynamic time warping algorithm has achieved 100% experimental accuracy,and the experiment based on BP neural network algorithm only took 0.000941 seconds in recognition while guaranteeing the accuracy of 99.4%,which reflects the efficiency and practicality of the method.After analyzing the feature of the contour points sequence of the hand gesture,a hand gesture recognition based on salient feature point selection is proposed.By setting a strategy,the redundant information points and some noise points on the gesture contour point sequence are deleted and the salient feature points are retained.The feature of hand gestures is extracted and represented based on those salient feature points.In the recognition period,a modified DTW algorithm is proposed.The absolute order of the salient feature point is used as an important parameter for the matching of two hand gestures.The data of the experiments show that this method is invariant to the rigid and articulated deformation and achieve good results in standard databases.Based on the depth information of hand images collected by Kinect sensor,a 3DSC based method is proposed for hand gesture recognition.The hand gesture contour is described in combination with the classical SC description and the depth information.After classifying gesture contour points into the bins of SC grid,the relative depth values of the contour points in the same bin are summed,and the result values are used as parameters of the grid instead of the number of contour points in the classical SC.The recognition based on this description method shows remarkable superiority in experiments and achieved high accuracy.