Based on a triad constant-rate biased RLG system, a mathematic model of the heading-sensitive error induced by calibration parameter variation errors in stationary initial alignment was set up. To tackle the problem that calibration parameters cannot be determined separately, installation misalignments were treated as a whole under the field test conditions, and then the heading-sensitive error coefficient was calibrated on-line using least square algorithm without external references. Utilizing the postulate system, field tests were done to estimate the on-line compensation accuracy of the heading-sensitive error coefficient by the on-line calibration method. The results of the experiments show that the on-line calibration and compensation method can eliminate the heading-sensitive error evidently and improve the yaw angle estimation precision in the initial alignment.