Activity recognition focuses on inferring current user activities by leveraging sensory data available. Nowadays, combining data driven with knowledge based methods has show an increasing interest. However, uncertainty of sensor data has not been tackled in previous hybrid models. To address this issue, in this paper we propose a new hybrid model to cope with the uncertain nature of sensors data. We fully implement the system and evaluate it using a large real-world dataset. Experimental results prove the high performance level of the proposal in terms of recognition rates