Gesture Recognition Technology Research Based On Kinect

Gesture recognition is a key technology for natural human-computer interaction,compared with the traditional mouse, keyboard and other mechanical equipment,gesture has the advantages of natural, intuitive and easy to understand, it is more closeto human’s daily communication habit. Kinect is equipment released by Microsoftwhich can capture RGB color image and depth image simultaneously. It can predicttwenty human skeleton joints’ three-dimensional coordinate from a single depth image,and it’s the ideal equipment for gesture recognition research based on computer vision.According to the recognized object, gesture recognition can be divided into static anddynamic gesture recognition. In this paper, static and dynamic gesture recognition isstudied respectively, using Kinect as an input device.Static gesture recognition classifies hand shape in single image, generally consistsof three steps: gesture segmentation, feature extraction and classification. The papercombines hand joint’s position and skin dectection method with adaptive threshold tosegment gesture, extracts Hu invariant moments and the number of fingers as thefeature, finally uses SVM to classify. Dynamic gesture recognition classifies hand’strajectory in continuous multi-frame images, generally consists of four steps: handsegmentation, hand tracking, feature extraction and classification. The paper getsskeleton joints’ position, uses skeleton joints’s trajectory as dynamic gesture’s feature,then uses distance-weighted dynamic time warping algorithm to calculate the distancebetween training sample and testing sample, finally uses K-NN to classify.The paper’s main research contents include the following three aspects:①Anew static gesture segmentation method is proposed. It combines hand joint’sposition and skin dectection method with adaptive threshold to segment static gesture.This method does not require a large number of training samples, can dynamicallyadjust the skin detection threshold according to the image captured in real time. Theexperimental results show that this method can overcome the interference of otherskin-alike region, and has perfect segmentation performence.②A new static gesture feature extraction method is proposed. It uses seven Huinvariant moments and the number of fingers to compose eight-dimensional featurevector as static gesture’s feature. Both Hu moments and the number of fingers areinvariant to rotation, translation and scale, the number of fingers can easily distinguish static gestures, but can’t uniquely represent static gestures. The experimental resultsshow that the average recognition rate of static gesture improves after adding thenumber of fingers on the basis of Hu invariant moments.③A new dynamic gesture recognition method is proposed. This paper presents anew distance-weighted DTW algorithm on the basis of dynamic time warping (DTW) tocalculate the distance between dynamic gestures’ training sample and testing sample,then uses K-NN classifer to classify. Because of each skeleton joint’s DTW diatance hasdifferent contribution to the final classification result, therefore, given different weightfor each skeleton joint’s DTW diatance. The experimental results show that theimproved algorithm has a higher average recognition rate.