This paper suggests a recursive least squares algorithm (BP-RLS) assisted by a BP neural network using three-axis geomagnetic sensor information to address the issue of missile-borne MEMS gyroscopes performance degrading in high overload environments, leading to changes in bias and scale factor. This algorithm achieves online calibration of error parameters for MEMS three-axis gyroscopes. Firstly, a linear online calibration model is established based on the cross-product integral of the geomagnetic vector and the condition of keeping the yaw angle constant. The recursive least squares algorithm is used to calculate the calibration parameters online, and the BP neural network is used to estimate the weight of the constraint equation in real-time. This guarantees that the error parameters will accurately converge when the noise condition varies. Simulation shows that the algorithm can achieve high-precision online calibration of MEMS gyroscope error parameters in a short period of time, meeting the accuracy requirements for measuring the attitude of rotating projectiles.