Discriminative Feature Fusion with Spectral Method for Human Action Recognition

摘要

In this paper, we propose an effective action recognition approach which differs significantly from previous interest points based approaches in that spectral information of video data is exploited. Firstly, we extract the motion interchange patterns feature and the HOG/HOF features of videos, respectively. We concatenate them into single feature representation. Secondly, Laplacian Eigenmaps is performed on the feature space to achieve the goal of dimensionality reduction. Spectral clustering is used to cluster the training set. Finally, SVM is taken for multi-class classification. Experiments using the UCF50 dataset and the YouTube dataset demonstrate that our approach achieve state-of-the-art performance.