A Novel Shallow Neural Network-Augmented Pose Estimator Based on Magneto-Inertial Sensors for Reference-Denied
Akos Odry1, Peter Sarcevic1, Giuseppe Carbone2
1Faculty of Engineering, University of Szeged, 6725 Szeged, Hungary.
Sensors (Basel, Switzerland)
|November 27, 2025
Summary
This study enhances mobile robot pose estimation by fusing Magnetic, Angular Rate, and Gravity (MARG) sensor data with shallow neural networks (NNs). The approach improves position/velocity accuracy in GPS-denied environments, crucial for quadcopter navigation.
Area of Science:
- Robotics
- Sensor Fusion
- Machine Learning
Background:
- Magnetic, Angular Rate, and Gravity (MARG) sensors are standard for mobile robot pose estimation but have limitations.
- Reliance on MARG data alone in environments lacking absolute references leads to significant estimation uncertainties.
- Accurate pose estimation is critical for autonomous navigation, especially in GPS-denied or landmark-deficient areas.
Purpose of the Study:
- To enhance the accuracy of position and velocity estimations by fusing MARG sensor data with shallow neural network (NN) models.
- To develop and train a cascade-forward NN for reliable estimation of true acceleration in dynamical systems.
- To evaluate the effectiveness of different NN topologies and training strategies for improved pose estimation.
Main Methods:
- Developed and trained three types of cascade-forward NNs for acceleration estimation using MARG measurements and signal features.
- Incorporated extended Kalman and gradient descent orientation filters during NN training.
- Conducted experimental validation using a low-cost flying quadcopter with a six degrees of freedom (6DOF) motion capture system.
Main Results:
- The proposed NN approach significantly improved the accuracy of rotation matrix-based acceleration vector calculation (Pearson correlation coefficient of 0.88 vs. 0.73 for baseline).
- Achieved reliable position/velocity estimations for quadcopter maneuvers within 10-second intervals (50m flight), with position errors between 2-4 meters.
- Demonstrated superior performance compared to baseline methods in enhancing estimation accuracy.
Conclusions:
- The developed NN-based fusion of MARG data offers a feasible and effective solution for enhancing pose estimation accuracy.
- The approach is particularly valuable for mobile robots operating in GPS-denied or landmark-deficient environments.
- Future research will focus on extending the application of this method to unknown environments.
Related Concept Videos
01:03Inertial Frames of Reference
8.6K
Newton’s first law is usually considered to be a statement about reference frames. It provides a method for identifying a special type of reference frame: the inertial reference frame. In principle, we can make the net force on a body zero. If its velocity relative to a given frame is constant, then that frame is said to be inertial. So, by definition, an inertial reference frame is a reference frame where Newton's first law holds valid. Newton's first law applies to objects with...
8.6K
01:27Non-inertial Frames of Reference
7.0K
A reference frame accelerating or decelerating relative to an inertial frame is a non-inertial frame. To help understand this, consider what taking off in an airplane, turning a corner in a car, riding a merry-go-round, and the circular motion of a tropical cyclone all have in common. All these systems are accelerating, decelerating, or rotating relative to the Earth; hence, they all are non-inertial frames. All these systems exhibit inertial forces, which merely seem to arise from motion,...
7.0K