In this paper, we use data from the Microsoft Kinect sensor that processes the captured image of a person, thus, reducing the number of data in just joints on each frame. Then, we propose a creation of an image from all the frames removed from the movement, which facilitates training in a convolutional neural network. Finally, we trained a CNN using two different forms of training: combined training and individual training using the MSRC-12 dataset. Thus, the trained network obtained an accuracy rate of 86.67% in combined training and 90.78% of accuracy rate in the individual training, which is a very good performance compared to related works. This demonstrates that networks based on convolutional networks can be effective for the recognition of human actions using joints.