Research And Implementation Of Gesture Recognition Based On Kinect

With the rapid development of virtual reality,artificial intelligence,multimedia and other scientific technologies,the means of human-computer interaction present the character of diversity.The research of gesture interaction method is one of the hottest topics.At present,gesture recognition technology has just started,and further research is still in need to be carried out to improve the real-time performance and accuracy.In this thesis,we mainly explored the design of gesture recognition system,focusing on the studies and improvements of fingertip detection algorithm and dynamic gesture recognition algorithm.Eventually,a test system was built to demonstrate and analyze the algorithms.The main contents of this thesis are as follows:(1)The common gesture segmentation and tracking algorithm are studied in the research,in which the process of gesture segmentation was analyzed,and the effect of segmentation was compared.Finally,the Kinect was chosen to be used to carry out the collection of depth images,hand positioning,image preprocessing and eight neighborhood tracking.A series of algorithms were used to divide the gestures.Based on the tracking algorithm,the drawbacks of the Meanshift algorithm and Camshift algorithm were analyzed,and it's proved that the Camshift algorithm is easily to lose the target when it's disturbed by occlusion.So,a Kalman filter prediction mechanism was introduced on the Camshift algorithm,in order to track and forecast gestures.(2)An improved algorithm for fingertip detection was provided,which verifies the deficiency of fingertip detection algorithm which is based on K-curvature and convex hull.In the end,a method which combined K-curvature and convex hull was used to identify the fingertip.The general gesture can be easily identified in this way.But when the gesture of fingertips was too close,it couldn't be identified correctly.In order to improve the fingertip close detection accuracy,an algorithm called "scan fitting method" was presented,and experiment to test its recognition rate.Based on the fingertip detection,we had the fingertip sorted and before and after frame matched algorithms are applied to calibrate the gesture whose initial state was known,and the trajectory of the index finger is tracked as well.(3)The dynamic gesture recognition algorithm was used to analyze the hand shape and trajectory characteristics of dynamic gesture.Lastly,Hu invariant was chosen as the character of hand shape,and angle vector was used as the trajectory feature.Then,based on the DTW algorithm,an efficient algorithm ET-DTW algorithm was proposed to optimize the recognition efficiency while improving the recognition efficiency.At the same time,the candidate sequence was sorted by LB_Keogh,and the algorithm was further optimized.Then the gesture recognition interactive system is designed for test.Experimental results show that the recognition rate increases from 90.2% to 94.1%,and the average detected time has decreased 0.091 s.