Detecting locomotion activities are critical for the analysis of human daily activities. In this paper, an adaptive on-line classification method is proposed to detect four lower limb locomotion activities -walking, running, stair ascent and descent -from the signals of a unique wearable sensor. The method is based on a non-parametric triplet Markov model, to detect gait phases and activities simultaneously, in an unsupervised way. This capability allows the model to work at run-time, and so to be used on-line. Also, an algorithm that adapts model parameters suits for a wide range of healthy human is presented. From this adjustment ability, an initial model can gradually approach to the dedicated activity patterns. Experimental results with ten healthy subjects show that our algorithm can reach an overall classification accuracy up to 99.20%, after the stabilization of parameters adjustment, regardless of the users' gender, height, activity speed . . . Overall, the proposed algorithm presents a good performance in on-line parameters learning and high accuracy in classifying lower limb locomotion activities from a fount-mounted inertial measurement unit-based wearable sensor