IMU Drift Correction via Neural Network Position Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Inertial measurement unit (IMU)-based position tracking in artificial reality systems suffers from accumulated error, known as 'drift error,' which causes the determined position to deviate significantly from the actual position over time, necessitating frequent corrections from secondary sensor systems.
Innovation Solution
A system utilizing inertial measurement units (IMUs) on multiple components of a user's body, such as a headset and handheld controller, feeds data into a trained neural network to predict relative positions, accounting for physiological constraints and reducing the need for continuous corrections from secondary sensors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If IMU-based position tracking is used, then the system can determine positions of body portions over time, but accumulated drift error causes the determined position to deviate from actual position
Solution Approach 1:
The system uses machine learning models trained on ground truth data to continuously correct IMU measurements. The model receives IMU data and compares predicted positions against known constraints and patterns, providing feedback to compensate for drift error and maintain accurate position tracking over time.
Solution Approach 2:
The patent replaces traditional mechanical integration methods (double integration of acceleration to get position) with a machine learning-based prediction system. The neural network learns physiological movement patterns and uses them to predict positions, substituting the error-prone mechanical integration process with a data-driven approach that accounts for human movement constraints.
2Duration of action of moving object
If continual integration of IMU measurements is performed to calculate position, then position can be determined over time, but measurement errors accumulate causing drift error
Solution Approach 1:
The system performs preliminary training of machine learning models using ground truth position data before actual tracking. This pre-training establishes accurate prediction capabilities that can be applied during extended tracking sessions, allowing the system to maintain precision over long durations without accumulating drift error through continual integration.
Solution Approach 2:
The patent changes the fundamental parameter for position calculation from integrated IMU measurements to machine learning predictions based on physiological patterns. By transforming the approach from mathematical integration to pattern recognition, the system maintains accuracy over extended periods without the quadratic error accumulation inherent in double integration.
3Reliability
If a neural network model trained on human movements is used, then physiological constraints are accounted for reducing drift error, but the system complexity increases
Solution Approach 1:
The system creates a computational model (neural network) that copies and learns from ground truth movement patterns. Instead of using complex hardware corrections, the patent uses software-based pattern copying where the network learns typical human movements and uses this knowledge to predict positions, reducing drift without adding physical complexity.
Solution Approach 2:
The machine learning model serves itself by learning from training data and then autonomously correcting IMU drift during operation. Once trained, the system uses its own learned physiological patterns to compensate for errors, reducing the need for external correction systems or complex hardware modifications.
Data Source
AI summary
A system having at least a first component and a second component positioned at different locations on a user's body (e.g., on the user's head and held on the user's hand). Each component includes at least one inertial measurement unit (IMU) configured generate measurements indicating acceleration and angular rate data. The generated measurements of the IMUs are used with ground truth information indicating the positions of the first and second component to generate a set of training data to train a neural network configured to predict a relative position between the first and second components based on IMU measurements received over a predetermined time period. Because the neural network is trained based upon movements of a human user, the neural network model takes into account physiological constraints of the user in determining how the set of potential positions of the different components may change over time, reducing potential error.