Research On Algorithm Of Motion-sensing Recognition

With the development of human-computer interaction techniques, motion-sensingrecognition technology has become an important part in human-computer interfaceresearch. As the core of motion-sensing recognition technology, motion-sensinggesture recognition technology has been the most popular research direction for itsflexible way of operation, being closer to human thought patterns and habits.This paper firstly analyzes the research status of motion-sensing gesture, findingout those gestures which are similar in translation and rotation have a poor recognitionrate and putting forward position-scanning algorithm. The algorithm based on copingwith similar gestures establishes a database including five common used gestures andhas four parts of work including image pro-processing, gesture contour extraction，gesture feature extraction and recognition, using MATLAB to do simulationexperiment.This paper firstly defines five common used gestures, uses general camera toobtain gesture images in complex background and converts those captured gestureimages rfom RGB color space to YCbCr color space.In the gesture contour extraction, calculating the exact color value in Cb and Crfrom skin color histograms, getting a certain threshold atfer several experiments,processing the threshold and change images to binary images, using self-definedgestures detection method to distinguish the hand and the face region，extractinggesture contours.In the gesture feature section, firstly using Hu invariant moments to extract sevencharacteristic values of gestures, using self-defined position-scanning algorithm andimproved area ratio algorithm to extract features, acquiring two new sequences offeatures.In the gesture recognition section, using SVM and BPNN algorithm to make threeexperiments. The first is to train seven characteristic values which are got rfom Huinvariant moments in SVM, and get a poor recognition rate, only41.37%. The second is to train all three kinds of characteristic sequences in SVM, and get a fast and goodrecognition rate，up to97.13%. The last is to train all three kinds of characteristicsequences in BPNN，and get a recognition rate over90%，but it costs a great deal oftime and system resource.Experimental results show that Hu invariant moments have a poor effect in copingwith those gestures which are similar in translation and rotation. Position-scanningalgorithm put forward by this paper has effectively solved the problem. In this paper,there have innovative algorithms in gesture contours extraction and gesture featuresextraction, and puts forward the design of education assistant system. This paper notonly has a high practical value, but also lays a good foundation for the further work.