Discriminative Features Extraction And Effective Classification In Human Motion Analysis

Humans perform enormous number of actions continuously in daily life. We identify and understand these activities unconsciously while interacting and communicating with people and the environment. If the machines and computers could also be aware of human gestures as effectively as human beings, a new world would be unfolded, filled with a large number of applications to facilitate our daily life. These significant benefits for the society have provoked the research on machine-based gesture recognition, which has already shown some initial advantages in many applications. For example, gestures can be used as commands to control robots or computer programs instead of using standard input devices such as touch screens or mice.In this thesis, a framework for discriminative, effective and computationally efficient for feature extraction and recognition algorithms for human gesture recognition systems is proposed based on spatiotemporal feature from RGB-D data where single example is available for training. In general, the framework can be roughly divided into two parts, feature extraction and classification. Both have significant influence on the gesture recognition performance.1. Shi-Tomasi corner detector algorithm is used for detection of key interestpoint in motion region in an image at different level in an image pyramid ofRGB(converted into gray-scale image) and depth image. While detectinginterest point, it might detect some futile points, so Lucas-Kanade trackingand filtering method is applied. Using the Lucas-Kanade method, absolutevelocity of each interest point as different level in pyramid is calculated. Thenonly those points are selected which satisfy the motion constraints.2. Improved Gradient Location and Orientation Histogram(GLOH) featuredescriptor is applied to capture the description of robust key interest point.Improved GLOH reduced the original quantized 16 bins into 8 bins whichproduce 136 descriptor vectors. Recognition rates are nearly equal but thecomputation time is almost half in Improved GLOH.3. In order to learn the discriminative model, all the features extracted from thetraining samples are clustered with the k-means algorithm to learn a visualcodebook. Then a sparse coding method named simulation orthogonalmatching pursuit(SOMP) is applied to achieve descriptor coding which mapeach feature into a certain visual codeword represented by some linearcombination of a small number of codewords.4. Recognizing the gesture, we propose to learn and classify gesture based on k-nearest neighbors method classifier. For each video in a training dataset,generate all the local motion descriptor and annotate them with the associatedgesture. The entire generated database is clustered; the k-nearest neighborclassifier is used to classify gestures occurring in the test dataset. A video isclassified according to the amount of neighbors which have voted for a givengesture providing the likelihood of the recognition. The effectiveness and efficiency of proposed framework is demonstrated by recognizing the actions of Cha Learn gesture dataset(CGD 2011).