Video Classification with Multi-Channel Convolutional Neural Networks

In this paper, we examined different ways for classifying videos with two-dimensional Convolutional Neural Networks (CNNs). By calculating energy of optical flow between frames, we found that classification with CNNs by feeding frames with high energy of optical flow can outperform results in comparison with feeding consecutive frames with highly similar content. It is demontrated that the energy of optical flow has straight relationship with classification accuracy.

کلیدواژه ها:

Video Classification ، Human Action Recognition ، Deep Learning ، Convolutional Neural Network (CNN) ، Energy of Optical Flow