A Novel Shallow Neural Network-Augmented Pose Estimator Based on Magneto-Inertial Sensors for Reference-Denied Environments

Magnetic, angular rate, and gravity (MARG) sensor-based inference is the de facto standard for mobile robot pose estimation, yet its sensor limitations necessitate fusion with absolute references. In environments where such references are unavailable, the system must rely solely on the uncertain MARG-based inference, posing significant challenges due to the resulting estimation uncertainties. This paper addresses the challenge of enhancing the accuracy of position/velocity estimations based on the fusion of MARG sensor data with shallow neural network (NN) models. The proposed methodology develops and trains a feasible cascade-forward NN to reliably estimate the true acceleration of dynamical systems. Three types of NNs are developed for acceleration estimation. The effectiveness of each topology is comprehensively evaluated in terms of input combinations of MARG measurements and signal features, number of hidden layers, and number of neurons. The proposed approach also incorporates extended Kalman and gradient descent orientation filters during the training process to further improve estimation effectiveness. Experimental validation is conducted through a case study on position/velocity estimation for a low-cost flying quadcopter. This process utilizes a comprehensive database of random dynamic flight maneuvers captured and processed in an experimental test environment with six degrees of freedom (6DOF), where both raw MARG measurements and ground truth data (three positions and three orientations) of system states are recorded. The proposed approach significantly enhances the accuracy in calculating the rotation matrix-based acceleration vector. The Pearson correlation coefficient reaches 0.88 compared to the reference acceleration, surpassing 0.73 for the baseline method. This enhancement ensures reliable position/velocity estimations even during typical quadcopter maneuvers within 10-s timeframes (flying 50 m), with a position error margin ranging between 2 to 4 m when evaluated across a diverse set of representative quadcopter maneuvers. The findings validate the engineering feasibility and effectiveness of the proposed approach for pose estimation in GPS-denied or landmark-deficient environments, while its application in unknown environments constitutes the main future research direction.