Abstract
Action recognition has vital significance for computer vision. Recently, deep learning has made breakthrough progress in action recognition. However, as two important branches of deep learning, Two-Stream relies on optical flow with complex computation and 3D convolution network is difficult for training. In this paper, we propose a novel Three-Stream Convolution networks after feature extraction for action recognition. For feature, we introduce three input features: RGB images, background subtraction feature with low complexity and historical contour feature. In order to optimize the discriminability of long-term actions, the historical contour feature is superimposed by background subtraction feature. For network structure, we present a convolution network stream for each feature input: RGB net, background subtraction sequence net and historical contour net. Finally, we merge three streams into one network with automatic network learning mechanism for action recognition to obtain a better recognition performance. We conduct experiments on two large main action recognition datasets UCF-101 and HMDB-51. Comparing the mainstream methods, the results verify the accuracy and high efficiency of our framework.