A 2D Convolutional Neural Network Approach for Human Action Recognition

Nowadays, deep neural networks are widely used for human action recognition (HAR) due to their ability to operate directly on the raw video inputs by extracting both the spatial and temporal information. Although the 3D convolutional neural networks as deep models have achieved superior performance, they remain computational expensive. In this paper we propose a 2D-CNN approach that learns robust feature representation from temporal information embedded into the motion history images of action videos. The proposed approach is simple and reduces the computational complexity imposed by the 3D-CNN approaches. The KTH database is used to validate our approach and the achieved results are compared favorably against the handcrafted state-of-the-art methods.