Human activity recognition systems based on video data have many functionalities and a wide range of possible applications. The classification of action types is an example. An action is a combination of gestures lasting a few seconds, such as waving or bending, and can correspond to physical exercises, such as raising the arms or leaning forward. Using information about the shape and movement of the silhouette in the foreground, it is possible to indicate the type of action performed by the human in the video sequence. This paper proposes the use of shape descriptors to create an action representation and a feed-forward neural network with a single hidden layer as a classifier. This approach can be identified as hybrid, as it combines classical machine learning methods with artificial neural networks. Research shows that this approach, despite its simplicity, produces comparable or even better results than more complex solutions based entirely on deep neural networks.