Human activity recognition is an active research area, especially in ambient assisted living environments. In such environments, residents' data are collected from sensors to be interpreted as human activities. The main constraint is that these activities have to be detected online and in real time for a continuous recognition. One major issue that remains a challenge to achieve is data segmentation. Usually, in the literature, the segmentation is either performed by following a fixed or a dynamic time-window length. As stated in several works, static time-window length has several drawbacks while adjusting dynamically the window length is more appropriate. However, most of the previous methods for dynamic data segmentation are based on two strong assumptions: the user's routine does not change and a pre-segmented data set can be provided for learning the time-window size. Yet, these constraints are not always verified. In this paper, we propose a novel method, DataSeg, that dynamically adapts the time-window size. DataSeg does not require pre-segmented data and it can be applied to different user routines. This is achieved by combining statistical learning and semantic interpretation to analyze the incoming sensor data and choose the better time-window size. The presented approach has been implemented and evaluated in several experiments using the real data set Aruba from the CASAS project. The experiments show the viability of the proposal.