Indoor positioning system based on MEMS (Micro-electromechanical Systems) technology is really self-measurement technology which does not depend on any external signal and other infrastructure. Inertial/magnetic sensors based on MEMS technology are typically composed of three orthogonal gyroscopes, three orthogonal accelerometers and three orthogonal magnetometers, which can form AHRS (attitude heading reference system) and pedestrian dead reckoning system. However, the inertial/magnetic sensors can offer good short term positioning in indoor environments, but the absolute position fixes of medium- to long- term have always been the research hotspot and difficulty. Therefore, a hierarchical calibration architecture based on inertial/magnetic sensors is proposed creatively, including the data layer-a axial-based temperature variation model is proposed to calibrate the raw data from MEMS sensors by the analysis and modeling of sensor error characteristics; the signal layer-a triple zero velocity update method is proposed to update integral initial value according to the pedestrians' gait characteristic; the information layer-an intelligent information fusion method is proposed based on machine learning to calibrate the orientation and positioning. Besides, a 3-D indoor positioning platform is set up based on MPU9250, and the effectiveness of the proposed hierarchical calibration architecture is verified based on the platform.