Study On Hand Gesture Recognition Used For Air Conditioning Control

As one of key technologies in human-computer interaction, hand gesture recognition attracts more and more researchers’ attention. Hand gesture recognition based on vision is natural, convenient and feasible without wearing sensors or some devices. Recently, hand gesture recognition has appeared in intelligent household and motion sensing games. The technologies not only provide convenience but also bring pleasures to life. However, there are some problems not solved successfully, such as low recognition rate, so they can not substitute for remote controls, mouse and other traditional devices completely. In addition, due to their high cost, hand gesture recognition are not been widely promoted. Hence, TK-C1480 BEC CCD camera,produced by Victor Company of Japan, is used to shoot hand gesture images and related issues emerged in real application are discussed and studied. Main research work is as follows:1) Image collecting: More than 2000 pictures with 12 kind of hand gestures under different illumination, weather and times are collected with actual scene.2) Hand gesture recognition: Common color spaces and skin detection models are analyzed, methods of elliptical model, simple threshold model combined with single Gaussian model and support vector machines based on pixels’ 8-neighbors are used for skin detection, by comparing their detection results, simple threshold model combined with single Gaussian is chosen.3) Feature extraction and recognition: Features for hand gesture recognition are studied, improved HOG and improved LSS descriptors are proposed.Due to local features extracted from hand gesture includes background information in real environment, an improved HOG combing skin similarity is proposed. Classification methods of SVM, SRC and NN are employed to classify hand gesture images with hand gesture features of HOG, SCHOG and improved HOG. Experiments are done on images of 12 kind of hand gesture and images of 6 kind of hand gesture in Marcel databases. Results show that, no matter which classification method is employed, hand gesture features of improved HOG can achieve higher recognition rate than HOG and SCHOG. SVM performs better than SRC and NN with same hand gesture features.According to characteristics of skin’s distribution in YCbCr color space, improved LSS descriptors are proposed for hand gesture recognition. SVM, SRC and NN are used for classification with hand gesture of LSS and improved LSS. Experiments are done on images of 12 kind of hand gesture and images of 6 kind of hand gesture in Marcel databases. Results indicate that, with same classification method, improved LSS can get higher recognition rate than LSS. With same hand gesture feature, SVM performs better than other two classification methods.4) Hand gesture recognition scheme design: A hand gesture recognition scheme is designed for a single image, moreover, a scheme for image sequences is also designed for a predefined hand gesture regulation.