Human Action Recognition Using Modified Bag of Visual Word based on Spectral Perception

Human action recognition has a very vast application such as security, patient care, etc. Background cluttering, appearance change due to variation in viewpoint and occlusion are the prominent hurdles that can reduce the recognition rate significantly. Methodologies based on Bag-of-visual-words are very popular because they do not require accurate background subtraction. But the main disadvantage with these methods is that they do not retain the geometrical structural information of the clusters that they form. As a result, they show intra-class mismatching. Furthermore, these methods are very sensitive to noise. Addition of noise in the cluster also results in the misclassification of the action. To overcome these problems we proposed a new approach based on modified Bag-of-visual-word. Proposed methodology retains the geometrical structural information of the cluster based on the calculation of contextual distance among the points of the cluster. Normally contextual distance based on Euclidean measure cannot deal with the noise but in the proposed methodology contextual distance is calculated on the basis of a difference between the contributions of cluster points to maintain its geometrical structure. Later directed graphs of all clusters are formed and these directed graphs are described by the Laplacian. Then the feature vectors representing Laplacian are fed to the Radial Basis Function based Support Vector Machine (RBF-SVM) classifier.