A common assumption in activity recognition is that the system remains unchanged between its design and its posterior operation. However, many factors can affect the data distribution between two different experimental sessions including sensor displacement (e.g. due to replacement or slippage), and lead to changes in the classification performance. We propose an unsupervised adaptive classifier that calibrates itself to be robust against changes in the sensor location. It assumes that these changes are mainly reflected in shifts in the feature distributions and uses an online version of expectation-maximisation to estimate those shifts. We tested the method on a synthetic dataset as well as on two activity recognition datasets modeling sensor displacement. Results show that the proposed adaptive algorithm is robust against shift in the feature space due to sensor displacement.