The Research Of Vision-based Hand Gesture Detection And Recognition Algorithm In Complex Background

The technology of hand gesture recognition based on vision is a very important researchtrend in the natural human-machine interaction. It has a broad application prospects in manyfields, such as the virtual (augmented) reality, robot control, intelligent appliances, the gamecontrol, as well as human-computer sign language interpreter. But the complex environmentand diversified, random hand gestures make hand gesture recognition based on vision verydifficult. So this article focuses on the hand gesture detection and recognition methods in thecomplex background.For hand gesture detection, the method combining face excluding, skin-colorsegmentation and hand shape comparison is selected in the paper. The skin-colorsegmentation and hand shape comparison methods are susceptible to the interference of face.Firstly, the Voila&Jones method and the Camshift algorithm is used to detect and track facein paper, and then the black circles or rectangles that is greater than the facial area are used toreplace and exclude face. During skin segmentation, an improved adaptive skin-color modelis proposed. The model combined the threshold method in YCrCb and the histogram methodin HSI. Since the color information of the model is obtained from the area of the face in realtime, so it has a good robustness for the illumination changing, the different users and noise.Finally, the hand shape comparing method based on Hu moments is adopted to locate handareas from the skin segmentation.For hand gesture recognition, combining the Hu moments and Bag-of-Features based onSpeeded Up Robust Feature (BoF-SURF) is proposed in paper. At the same time, thealgorithm combining the integration of features and multi-class support vector machine(SVM)are adopted. The fusion feature can represent gesture from hand shape and internal key points.It has a good adaptability for the gesture changing in scale, light, rotation and partialocclusion. Taking into account the two information has a different influence for gesturerecognition, so different weights are set, and its best values are determined by experiment.The result shows that the fusion feature has a good performance. Its’ recognition accuracy ishigher than the separate feature. The96.33%recognition accuracy can be achieved byselecting appropriate weights, The proposed method shows better real time performance when compared to others with processing time.