Multiple-resident activity recognition is a major challenge for building a smart-home system. In this paper, conditional random fields (CRFs) are chosen as our activity recognition models for overcoming this challenge. We evaluate our proposed approach with several strategies, including conditional random field with iterative inference and the one with decomposition inference, to enhance the commonly used CRFs so that they can be applied to a multipleresident environment. We use the multi-resident CASAS data collected at WSU (Washington State University) to validate these strategies. The results show that data association of non-obstructive sensor data is of vital importance to improve the performance of activity recognition in a multiple-resident environment. Furthermore, the study also suggests that human interaction be taken into consideration for further accuracy improvement.