An Efficient Action Recognition Framework Based on ELM and 3D CNN

Deep neural network is shown to be the most efficient method for video representation and has achieved state-of-art results on different datasets of action recognition. In this paper, we proposed a hybrid architecture which integrates deep convolutional neural networks and extreme learning machine. The hybrid structure makes the most of their advantages: in the first stage the deep residual 3D network learns the features from both temporal and spatial sequences, then the ELM, instead of traditional classifiers, classifies the actions without tuning the parameters. The resulting network can not only extract the representation fully, but also obtain more accurate results faster. We show the effectiveness and outperformance of the proposed strategy on experiments.