Radar Kinematic Data Augmentation for Robust Gesture Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine-learning models for radar-based gesture classification suffer from varying quality and poor accuracy due to covariate shift, where training datasets are not representative of deployment scenarios, leading to incorrect estimations.
Innovation Solution
A method of populating a training dataset by applying transformations to time-resolved kinematic observables extracted from radar measurement data, tailored to match the deployment configuration, using data augmentation techniques to generate synthetic input samples.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a machine-learning model is trained on standard radar measurement data, then the model can perform gesture classification, but the accuracy is poor due to covariate shift between training and deployment scenarios
Solution Approach 1:
The patent applies parameter changes by transforming the time-resolved values of kinematic observables through various operations (time-scaling, amplitude-scaling, noise-injection, shifting) to create augmented training samples that reflect different deployment conditions. This resolves the covariate shift problem by ensuring the training data encompasses the variability expected in deployment scenarios, thereby improving both accuracy and reliability.
Solution Approach 2:
The patent creates synthetic copies of existing radar measurement samples by applying transformations to generate augmented training data. Instead of collecting extensive real-world data from multiple deployment scenarios, the method copies and transforms existing samples to simulate various deployment conditions, effectively resolving the generalization problem without requiring additional physical measurements.
2Measurement precision
If transformations are applied to generate augmented training samples, then the accuracy improves, but the data processing complexity increases
Solution Approach 1:
The patent applies preliminary action by performing data augmentation during the training data preparation phase rather than during real-time inference. All transformations (time-scaling, amplitude-scaling, noise-injection, shifting) are pre-applied to create the training dataset, which eliminates the need for complex processing during deployment and resolves the contradiction between improved accuracy and processing complexity.
3Measurement precision
If more diverse training data is collected to match deployment scenarios, then the model accuracy improves, but the time and resources required for data collection increase
Solution Approach 1:
The patent uses copying to create synthetic diverse training data from a limited set of actual measurements. By applying transformations to existing samples, the method generates virtual data that mimics diverse deployment conditions without requiring physical collection of data from multiple scenarios, thereby resolving the contradiction between accuracy and data collection time.
Solution Approach 2:
The patent applies parameter changes to existing training samples through transformations that modify time-resolved values, creating synthetic data with varied characteristics. This approach generates diverse training data by altering parameters of existing samples rather than collecting new data, effectively resolving the contradiction between improving accuracy through data diversity and reducing data collection time.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Various examples of the disclosure generally pertain to providing estimations such as class estimations, e.g., gesture class estimations, based on radar measurements and using a machine-learning model. Various examples specifically relate aspects associated with the training of such a machine-learning model. Data augmentation techniques are for populating a training dataset. Data augmentation is executed for values of object-specific observables such as range (601) or velocity (604).