The development of the computing power and sensing ability of mobile devices allows them to provide various contextadapted services to users. The on-body position of mobile devices, which is one kind of important context information, affects the recognition of other human activities and the adaptability of many mobile applications. The study provides a method to recognize the on-body positions of mobile devices, inspired by the analysis that different positions on the body have distinguishable rotation patterns. The research then fuses the data sensed by the accelerometer and the gyroscope to calculate the rotation radius, the magnitude of the angular velocity as well as the gravity acceleration and then extract a set of features. A classifier based on the Random Forest is used for classification and compared with the solution based on the support vector machine. To evaluate the method, the paper conducts an experiment on a public dataset with 3 types of positions and 13 types of activities. Results show that the method achieved an accuracy of 95.39% on average in the cross validation and indicate that when rotation is the main component in the movement and the direction of the gravity acceleration is stable, the information about rotation variation and the ensemble classifier are useful to improve the classification accuracy. Compared to previous works, it is able to classify the positions more precisely and has more generalization ability for new users and new activities.