Research For Gesture Recognition In Complex Environment

Vision based gesture recognition is an important issue in the field of human-computer interaction. However the actual scene in the light effects such as dramatic changes, complex background, how to design efficient and robust gesture recognition algorithm is a difficult problem. This paper studies the gesture recognition method under complex background.Gestures segmentation is the most key step of gesture recognition. Based on valve of classical skin segmentation model on complex environment is not robust, but also based on Gaussian model of skin segmentation method time complexity is high. Therefore, it is hard to reach real-time requirements. This paper proposed an algorithm way,combining of YCbCr and HSV two kinds of skin segmentation methods get the skin regional,however it's not easy to distinguish similar skin regions. So background difference of movement segmentation method is added on to filter out similar skin regional, Get a rough outline of hand contour. Using fusion of gradient histogram (HOG) and the Hu moment for feature extraction, the invariance of Hu moment in rotation, translation and scale make up HOG feature which is not sensitivity to rotation. The gesture area will eventually be put in the trained classifier to identify and get a good result.There needs a certain range of movement when hand gesture recognize above in real-time detection. In some complex environments segmentation result is not very ideal, artificial selection of feature extraction methods for larger gestures deformation and rotation is not reflected a strong ability to identify and use of deep learning method for end-to-end learning capacity, Faster R-CNN-based gesture recognition method proposed in this paper. Using Convolutional neural networks in the region,with the 7-layer network framework we fine tuning network parameters in the framework of faster R-CNN model using my 8 categories gestures database.Adjust mini-batch size proportion of positive and negative samples, base_size,candidate box area number after nms successfully detect smaller targets gestures.Experiments show that the Faster R-CNN-based gesture recognition methods for the real-time detection are very robust.