Behavior recognition through ambient assisted living solutions for elderly people, represents an ambitious challenge for actimetry. Numerous and versatile solutions have been deployed. However, a commercial adoption is still pending, due to scalability and acceptability constraints. Most researches in ambient assisted living appear to have an heavy design, where precise features are first selected, and hardware architecture is designed accordingly. Although it may provide interesting results, such approach leads to a lack of scalability. This is why we experimented a lighter approach for a real deployment. The complexity is shifted from hardware to software, and we aim to make meaningful informations emerge from simple and generic sensor data, in order to recognize abnormal and dangerous situations. In this paper, we will describe how to retrieve consistent informations, so that resident's behavior may be observed, and that a light and generic approach fits in large scale deployments, with acceptable cost and scalability.