To ensure the high precision output of micro-electro-mechanical system (MEMS) accelerometers, the real-time compensation calibration needs to be implemented in the application process. In this study, an acceleration self-calibration model between the measured value and the real value was established, and multiple sets of static observation samples of the accelerometer at different positions were filtered. The Levenberg-Marquardt (LM) algorithm was combined with the least squares method to calculate the model parameters, which solved the problem of initial value dependence of the LM algorithm. For the static output of the accelerometer at different positions, the posture data that can be used for the least squares method were selected after filtering, which were applied to modify part or all the kth iteration model parameters, being the initial value of the (k+1)th iteration; other posture data were used to train the (k+1)th model parameters by the LM algorithm, realizing the closed-loop and real-time calibration of the accelerometer in application. Taking the application of intelligent insoles as an example, the calibration results of traditional 12-position method,ellipsoid fitting method, LM algorithm, and LM & least squares method were compared. Experimental results show that in the long-term use of intelligent insoles, the proposed LM & least squares method could eliminate the inaccurate calculation of model parameters due to the setting of the initial value of LM algorithm. Besides, it could realize the real-time acquisition, calculation, and calibration of the target, and achieve the precision of the same magnitude as the traditional calibration methods.