Human action recognition based on sub-data learning

Abstract

Human action recognizing nowadays plays a key role in varieties of computer vision applications while at the same time it’s quite challenging for the requirement of accuracy and robustness. Most current computer vision methods focus on algorithms designing classifiers with handcrafted features which are complex and inflexible. To automatically extract both spatial and temporal features, in this paper we propose a method of human action recognition based on sub-data learning which combines the proposed 3D convolutional neural network (3DCNN) with the One-versus-One (OvO) algorithm. We also employ effective data augmentation to reduce overfitting. We evaluate our method on the KTH and UCF Sports dataset and achieve promising results.