Real- time gesture recognition using 3D images

Abstract

KINECT has been recently introduced in the market as a low cost 3D acquisition device, so it will be interesting to discover the power of this device when we use it for gesture recognition. In this thesis, we propose a real-time gesture recognition system using 3D sensor that transforms gestures into a set of useful words using different machine learning algorithms and taking into consideration temporal features. A depth image, which is provided by KINECT, will be used to construct a skeleton of the human body. We have used Nearest Neighbor (NN) with different distance formulas, Self Organizing Map (SOM) and Hidden Markov Model (HMM) for recognition. The result of this thesis using the 10 fold cross validation shows that HMM may provide recognition accuracy up to 96 percent, while using NN algorithm with Spearman distance we can obtain around 90 percent accuracy and around 75 percent of accuracy using the SOM algorithm. All three algorithms has works in real-time