Abstract
Human action recognition is an essential topic in computer vision research that has received a wide concern. In the previous decade, deep Convolutional Neural Networks (CNN) have offered an effective execution for several action recognition datasets. However, utilizing an immense amount of training data and using backpropagation algorithm in the training stage have a negative influence on its efficiency for action recognition problem. To help alleviate these limitations, we present a novel technique for human action recognition based on Principal Component Analysis Network (PCANet), which is a simple deep learning network. A subset of frames is selected from each action while for each frame a feature vector is calculated from the previously trained PCANet. All feature vectors are then fused and their dimensionality are reduced using Whitening Principal Component Analysis algorithm (WPCA). Finally, Support Vector Machines (SVM) classifier is employed for action recognition. We assess the proposed approach on the challenging dataset, KTH Human Action Dataset. Our experimental results using the leave-one-out evaluation strategy show the efficiency of the proposed method.