Improved GLOH approach for one-shot learning human gesture recognition

Abstract

A method is presented for One-Shot Learning Human Gesture Recognition. Shi-Tomasi corner detector and sparse optical flow are used to quickly detect and track robust key-points around motion patterns in scale space. Then Improved Gradient Location and Orientation Histogram feature descriptor is applied to capture the description of robust key interest point. All the extracted features from the training samples are clustered with the k-means algorithm to learn a visual codebook. Subsequently, simulation orthogonal matching pursuit is applied to achieve descriptor coding which map each feature into a certain visual codeword. K-NN classifier is used to recognizing the gesture. The proposed approach has been evaluated on ChaLearn gesture database.