The fall detection algorithm, which can recognize the fall of human body by collecting the acceleration signals in different directions of the body, is an important part of fall detection system for the elderly. The system, however, may have errors during analyzing the acceleration signal, due to that the coordinate system of the transducer does not coincide with the one of human motion. Furthermore, voltage variation of the battery also influences the accuracy of the acceleration signal. Therefore, in this paper, a fall detection system based on the 3D-acceleration transducer MMA7260 is designed, which can calibrate the acceleration data through compensation of voltage and transformation of coordinates. Experiments illustrated that the proposed method can accurately transform the collected data from the coordinate system of the transducer to that of the human motion, and can recognize various postural changes in the course of the motion of human body.