Real-time ellipse detection is an important yet challenging task, since the estimation of the five parameters of an ellipse requires heavy computation. This task is even more challenging when the processing must be done on a mobile device with limited computational power. The typical trade-off between accuracy, efficiency and limited resources of embedded vision programming must be accounted. In this paper we present a novel strategy for edge point selection, which allows to drastically reduce the number of edge points to be evaluated for parameters estimation, making embedded mobile vision applications feasible. Extensive results show the increased efficiency of the proposed method over state-of-the-art ellipse detectors, in synthetic and challenging real images, and in a live mobile application.