Human Action Recognition Based on RGB-D and Local Interactive Regions Detection

摘要

We propose a novel method to recognize human actions by fusing information from RGB-D sensors. Human action recognition is a challenging task because of the complexity movements, the variety of actions performed by different subjects and the changes of view and illumination. We propose to detect human motion from body parts by extracting sets of spatial-temporal interest points from RGB sequence and projecting them into depth map. Then, extract local interactive regions as supplementary information for action recognition. An improve classifier based on linear SVM coupled with dynamic time warping is developed for classification. We evaluate our method on two public datasets, including MSRDailyActivity3D dataset and ReadingAct RGB-D action dataset.