Hand Gesture Recognition System

Hand Gesture Recognition is a well-researched topic in the community of Machine Learning, Computer Graphics and Image Processing. The system which are based on Recognition technology follow mathematically rich and complicated algorithms whose main aim is to teach a computer different gestures. Because there are very large sets of gestures, the number of methodologies to identify the set of gestures is also large. In this thesis, I have concentrated on the gestures are based on hands. The thesis is divided into two sections namely: Static mode and Dynamic. The Static mode concentrates on gestures based on still images and Dynamic mode concentrates on gestures based on image sequence. As every hand gesture recognition system, the recognition paths has been divided into basically four parts for Static mode: Segmentation, Feature Extraction, Feature Selection and Classification. In the static mode the algorithm used are the Graph Cut algorithm, Bacterial foraging optimization algorithm, Support vector machine, binary tree color quantization algorithm, block-based discrete cosine transform. The Graph Cut algorithm uses the min-cut of a graph to separate the non-hand pixels from the hand pixels. The min-cut of the graph is found out by the max-flow algorithm. The binary tree color quantization algorithm is used to cluster the pixels into required number of clusters. The BFO algorithm is used to find the optimum value of parameter that are either required to be maximized or minimized. The BFO is an evolutionary algorithm which is a reflection of the swamping behavior of the E. Coli bacteria. The algorithm is a non-linear form of optimization and the convergence of the algorithm is faster than the other evolutionary algorithms. For Dynamic mode the path has been divided into four parts: Segmentation, Tracking, Feature Extraction, Vector Quantization and Classification. The Dynamic mode uses 150 frames of image data to trace the path of the hand and finds the most likely gesture. The hand isolation is done by use of Gaussian Mixture model. To make the system as fast as possible the tracking of hand was more preferred to be fast than accurate. So some amount of accuracy was sacrifice for the sake of performance. As the sequence of image is involved the Hidden Markov model was preferred method for the classification. The training of the HMM was done by the method described by Baum- Welch which is the maximization of the expected value of the parameters of the HMM. The training was followed by the testing where an image sequence of 150 frames was passed to the system. The Viterbi algorithm was used for the testing purposes. The Viterbi algorithm finds the most like sequence of states for which that particular sequence of observation is taken out