Multiple Feature Voting based Human Interaction Recognition

Most of currently interaction recognition methods always need to segment the spatio-temporal features to the individuals involved in the interaction or need to build complex action models to present the human interaction. A novel method is proposed without considering the feature segmentation and complex action model in this paper. The proposed method utilizes two simple features i.e., improved BoW descriptor of interest points and HoG descriptor to respectively represent the local characteristics and global characteristics of human interactions. The classification voting histogram of BoW features and HoG characteristics are obtained by frame to frame nearest neighbor classifier respectively. Finally, recognition result is achieved by weighted fusing the classification voting histogram of these two feature. The method is tested on UT-Interaction dataset. Experiment result show that the method achieved the better recognition performance with simple implementation.