This paper presents a Markov Logic Networks (MLN) approach for the on-line recognition of human activities in a smart home. The method recognises activity from non visual and non wearable sensors data. The classification model benefit from a logic formal representation and uses probabilistic inference to deal with uncertainty. The evaluation was carried out on a real smart home where 21 participants performed several daily activities recorded by microphones and several classical home automation sensors. The MLN approach reaches an accuracy of 85.3% while the basleine support vector machine and naive Bayes approches leads to 59.6% and 66.1% respectively. Results show not only the great abilities of MLN as a classifier but also its potential to be integrable into a formal knowledge representation framework.