HAND GESTURE RECOGNITION USING CNN

Abstract

An intriguing area for the development of intelligent vision systems is the automatic 
identification of human gestures from camera pictures. We present a convolutional 
neural network (CNN) approach to detect human task-related hand motions in still 
images captured by a camera. In order to attain robustness performance, the CNN is 
trained and tested using data obtained from the skin model and the calibration of hand 
position and orientation. Our skin model is trained using a Gaussian Mixture model 
(GMM) to effectively exclude non-skin colours from images, which is necessary since 
light conditions have a significant impact on skin colour. The goal of the hand 
orientation and position calibration is to rotate and translate the hand picture to a 
neutral position. The CNN is then trained using the calibrated images. By conducting 
an experiment, we were able to validate the suggested approach for human gesture 
recognition, which yields consistent results regardless of lighting or hand orientation. 
This method's practicability and dependability were shown in our experimental 
assessment of seven people executing seven hand gestures with average recognition 
accuracies of 95.96%.