For the last several decades, Human Activity Recognition (HAR) has been an intriguing topic in the domain of artificial intelligence research, since it has applications in many areas, such as image and signal processing. Generally, every recognition system can be either an end-to-end system or including two phases: feature extraction and classification. In order to create an optimal HAR system that offers a better quality of classification prediction, in this paper we propose a new approach within two-phase recognition system paradigm. Probabilistic generative models, known as Deep Belief Networks (DBNs), are introduced. These DBNs comprise a series of Restricted Boltzmann Machines (RBMs) and are responsible for data reconstruction, feature construction and classification. We tested our approach on the KTH and UIUC human action datasets. The results obtained are very promising, with the recognition accuracy outperforming the recent state-of-the-art.