Gesture Recognition Model Training With Side-Independent IMU Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional gesture recognition models for wearable devices require separate training for each side of the user, leading to inefficient memory and processing usage since most devices are only used on one side, resulting in significant waste.
Innovation Solution
A transformation is applied to gesture data collected from a single side of the user to make the model independent of the side, allowing it to be used on either side, reducing memory and processing requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If separate gesture recognition models are trained for each side of the user, then gesture detection accuracy is maintained, but memory and processing resources are doubled
Solution Approach 1:
The patent applies asymmetry by collecting gesture data from only one side of the user body during training, then using transformation techniques to make the model applicable to both sides. This asymmetric data collection approach reduces the quantity of training data and model parameters needed, thereby reducing memory and processing resources while maintaining gesture detection accuracy through mathematical transformations that adapt the single-side model to handle both-side gestures.
Solution Approach 2:
The patent achieves universality by creating a single gesture recognition model that can function for both left and right sides of the user body. Through data transformation and normalization techniques, the model trained on one side becomes universally applicable to both sides, eliminating the need for separate models and reducing resource requirements while maintaining comprehensive gesture detection capability.
2Adaptability or versatility
If gesture data is collected from both sides of the user, then the model can recognize gestures on either side, but training complexity and data processing requirements increase
Solution Approach 1:
The patent inverts the conventional approach by not collecting data from both sides during training. Instead, it collects data from one side and then applies transformation techniques to make the model applicable to both sides. This inversion simplifies the training process and reduces data processing requirements while achieving the same adaptability goal.
Solution Approach 2:
The patent applies parameter changes by transforming the gesture data parameters through mathematical operations that normalize the data from one side to represent both sides. This parameter transformation approach maintains model versatility while reducing training complexity by working with a single set of training parameters rather than separate parameters for each side.
3Quantity of substance
If a single gesture recognition model is used for both sides, then resource usage is optimized, but gesture detection accuracy may decrease
Solution Approach 1:
The patent introduces transformation techniques as an intermediary between the single-side training data and the both-side application. This intermediary process involves mathematical transformations and normalizations that bridge the gap between asymmetric training data and symmetric application requirements, maintaining gesture detection accuracy while using a single optimized model.
Data Source
AI summary
In a method for training a gesture recognition model, gesture data is collected from an inertial measurement unit (IMU) positioned on one side of a user, wherein the IMU is capable of collecting data when positioned on either side of the user. A transformation is applied to the gesture data, wherein the transformation generates transformed gesture data that is independent of either side of the user. A gesture recognition model is trained using the transformed gesture data.