A lot of research has been done for human activity recognition. But most of it uses a static and immutable set of sensors known beforehand. This approach does not work when applied to a ubiquitous or mobile system, since we cannot know which sensors will be available in the users’ surroundings. This is why we consider here an opportunistic approach, where each sensor individually trained are able to bring its own knowledge. Inspired by the Opportunity project, we propose to evaluate both the effectiveness of using a Random Forest (RF) classifier to train the sensors and the robustness of fusing the results using a weighted majority vote. We found that RF gave better and more robust results than the other classifiers formally tested by Opportunity.