GPU Accelerated Real-Time Hand Posture Recognition

Abstract

With the development of computing current user interaction approaches with keyboard, mouse and pen are not sufficient. Hand postures can be used as an input device is very good method for natural Human Computer Interaction for many purposes. However tracking hand postures is a very challenging problem because of high degrees of freedom of human hand. This thesis proposes a new approach to solve this hand posture tracking in real time environment. Statistical analysis is used for this approach. This is done in low-level hand posture tracking using with Haar-like features and the AdaBoost algorithm. This approach is capable of processing images extremely rapidly while achieving high detection rates. The Haar-like features can efficiently catch the feature exaction properties and the AdaBoost learning algorithm can speed up the performances by making strong cascade classifiers. To recognize the hand postures image is scan by a final cascade classifier. In this thesis, further we develop a GPU accelerated real-time and robust hand posture recognition. We use Compute Unified Device Architecture (CUDA), a C based programming model from NVIDIA for and develop optimized parallel implementations of the Viola and Jones algorithm. For speed up the hand posture tracking we optimize the integral image calculation in the algorithm using threads in the Graphical Processing Unit.We evaluate our hand posture recognition system using both static images as well as using live frames captured from an A4TECH high quality web camera under realistic conditions. Our experimental results shows that our optimized hand posture recognition system achieve much greater detection speeds and accuracy as compared to existing work.