Human contour extraction from RGBD camera for action recognition

Spatio-temporal cuboid pyramid (STCP) for action recognition using depth motion sequences [1] is influenced by depth camera error which leads the depth motion sequence (DMS) existing many kinds of noise, especially on the surface. It means that the dimension of DMS is awfully high and the feature for action recognition becomes less apparent. In this paper, we present an effective method to reduce noise, which is to segment foreground. We firstly segment and extract human contour in the color image using convolutional network model. Then, human contour is re-segmented utilizing depth information. Thirdly we project each frame of the segmented depth sequence onto three views. We finally extract features from cuboids and recognize human actions. The proposed approach is evaluated on three public benchmark datasets, i.e., UTKinect-Action Dataset, MSRActionPairs Dataset and 3D Online Action Dataset. Experimental results show that our method achieves state-of-the-art performance.