Classification of Video Dynamic Texture Based on Lie-SVM

Abstract

In order to more accurately measure the dynamic texture pattern of video clips, it is necessary to analyze the topological space of this pattern; to this end, we propose a method for classifying video clips based on dynamic texture features and a support vector machine (SVM) with a Lie group kernel. First, the autoregressive moving average model (ARMA) is established to describe the dynamic texture of video, which is mapped to the Lie group manifold to form the matrix shape of Gaussians (SOG) descriptors. Second, the distance measurement criterion on Lie group manifold is used as the kernel function of SVM, and the algorithm of multi-classifier with Lie-SVM is designed. Finally, the example verifies that the multi-classifier for dynamic texture features of flame video based on Lie-SVM has better recognition rate (accuracy improvement 15.26%) and lower time consumption (runtime reduction 2.3%), which provides a new idea for the recognition of dynamic image texture features. The results show that the average recognition rate of flame video dynamic texture based on Lie-SVM algorithm is 11% higher than that of Martin distance algorithm, and through the influence of the dimension of state sequence on the average recognition rate, it can be concluded that the optimal condition of Lie-SVM algorithm is the singular value decomposition of observation sequence with 12 dimension.