Abstract: This paper presents an automatic turns detection and annotation technique which works from unlabeled captured locomotion. Motion annotation is required by several motion capture editing techniques. Detection of turns is made difficult because of the oscillatory nature of the human locomotion. Our contribution is to address this problem by analyzing the trajectory of the center of mass of the human body into a velocity-curvature space representation. Our approach is based on experimental observations of carefully captured human motions. We demonstrate the efficiency and the accuracy of our approach.