We propose in this paper a novel architecture for human activity monitoring, following conceptual, technical and experimental claims. From a conceptual viewpoint, we propose to approach the interpretation of sensor data as embedded into a multidimensional frame involving functional and non-functional requirements. Functional requirements involve considering the monitored person's specicities as well as the task to be performed. Non-functional requirements qualify the system activity. This frame of interpretation is continuously rened, to cope with evolving situations or expectations from the Observer. From a technical viewpoint, we propose to develop a multi-Agent architecture as a means for dependable, exible monitoring. This paradigm allows to handle multiple, heterogeneous entities in a unied way. The Agents process incoming data with a dynamic population of hypotheses on several abstraction levels. This reasoning is abductive and fuzzy in nature. From the experimental viewpoint, we propose a dedicated evaluation approach to estimate the interpretative process unfolding. Functional and non-functional properties are presented to discuss the system's effectiveness, informativeness, sensitivity, eficiency and robustness, some of which are supported by qualitative, analytical discussions, others by quantitative measures.