An Improved Method for Interest Point Detection in Human Activity Video

Abstract

An improved method for human activity recognition in RGB-D video using improved color-depth local spatio-temporal features (CoDe4D) detection method is presented in this paper. Firstly, histogram equalization and correction function are employed to suppress noise of rgb frame and depth frame, respectively. Then a saliency map is constructed by the improved CoDe4D method. In this way, feature points can be obtained by the saliency map, which integrates both depth information and RGB information. Next, the depth cuboid similarity feature (DCSF) is utilized to describe feature vectors. Meanwhile, visual words are generated by bag of feature method. To further improved the estimation accuracy, the SVM with generalized histogram intersection kernel is applied to train and predict categories. It shows good performance on MSR Daily Activity 3D datasets.