AMI-IAGDA: A Robust IMU Orientation Estimation Framework With Quaternion-Based Magnetic Interference Detection and Adaptive Gradient Descent

IMU-based motion capture systems face challenges of attitude drift and magnetic interference. To address these, this paper proposes a hybrid algorithm named AMI-IAGDA, which integrates an Anti Magnetic Interference (AMI) model with an Improved Adaptive Gradient Descent Algorithm (IAGDA). To mitigate the impact of magnetic interference, an Anti Magnetic Interference model is developed. Magnetic anomalies are detected using a weighted k-nearest neighbor (WKNN) method based on a quaternion similarity metricS(q, p). The WKNN classifier takes raw 9-axis IMU data as input features and employs an inverse-distance weighting scheme to classify samples, enabling dynamic validation of magnetometer data and adaptive fusion mode switching. This approach significantly enhances stability and estimation accuracy under magnetically disturbed conditions. To address the issues of uncompensated gyroscope bias drift and convergence instability due to fixed step sizes in IMU attitude estimation, an adaptive gradient descent algorithm is proposed. By incorporating a proportional–integral (PI) controller into the Gradient Descent Algorithm (GDA) framework, the algorithm effectively suppresses drift errors caused by cumulative integration over time. In addition, a dynamic step size adjustment strategy is designed based on the directional cosine criterion, accelerating the convergence process and enabling adaptive tuning within the attitude optimization algorithm. The proposed AMI-IAGDA achieves both enhanced attitude estimation accuracy and faster convergence. Under intense motion with magnetic interference, experimental results show that, compared to the conventional GDA, the proposed algorithm reduces the Euler angle RMSE by 80.96%, 40.64% and 63.49%, respectively. Compared to mainstream anti-interference algorithms, the proposed method demonstrates lower Euler angle errors and faster convergence, validating its effectiveness and robustness.