Dynamic Hand Gesture Recognition Using Computer Vision and Neural Networks

Using hand gestures is one of the most natural ways of interacting with the computer and most importantly correct interpretation of moving hand gestures in real-time has many applications. In this paper, the author has designed and developed a system which can recognize gestures in front of a web camera real time using motion history images (MHI) and feedforward neural networks. Firstly, background from captured frames is removed using Gaussian mixture based background/foreground segmentation algorithm in order to capture moving areas in the frame and thereafter median filtering has applied to remove random noise from the frame. Then binary thresholding with Otsu's binarization has applied and it will identify optimal threshold value and these processed frames are merged and cumulative motion history image is generated using a developed algorithm based on the structural similarity measure. Structural similarity between the cumulated image and the initial frame also calculated and used in this algorithm. Finally feed forward neural network with stochastic gradient-based optimizer has used to classify the gestures.