Robust hand gestures tracking method in cluttered background based on multilayer perceptron
Mehdi Heidaryan, Fardad Farokhi · 2015
This paper presents the use of multilayer perceptron (MLP) neural network with supervised learning algorithm to track hand gestures. Two networks are used. The first network is trained to detect skin and the second identifies face, hand and its gesture. A binary image made from the first network is transmitted to second network as its input. Both networks have two output classes, first network: skin and non skin in the other network due to existence of hand and face in binary image, first class is gestures of hand and second class is face. Hand gestures are divided into 5 cases: hand closed, hand open, two fingers (victory), three fingers and Index and Little finger. The two networks are applied on video sequence frame to frame. Features in the second network are selected so that tracking is independent from size and rotation of hand and computation is decreased because of excising common terms between DFT and DCT. Results show both networks have high accuracy. Matlab is used and not claim real time processing (2 fps) but in our future work these two networks are implemented on FPGA in real time.