Many solutions have been proposed for 3D human pose estimation from video data. However, only a few of them take into account temporal features. In this article, we present a method focusing on this temporal aspect and show promising results. Our approach consists of two parts. The first one concerns the creation of a dataset that contains a variety of motion features. Based on this dataset, the second one deals with the training of a DCNN-based model, which takes as input the 2D pose estimations directly computed from videos. Here we present the first training tasks and results obtained using our deep neural network model to directly estimate 3D poses. Three models were trained using the same architecture applied on several configurations of our dataset. Using a small benchmark, we evaluate our network architecture.