Generative AI Augmented Image Data for Motion Sickness Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The limited number of training data sets available for machine learning models to detect kinetosis in vehicles, which hinders the development of effective models due to insufficient sensory equipment and limited participant studies, leading to suboptimal performance.
Innovation Solution
A method and device that utilize a multimodal training data set with kinetosis state as ground truth, focusing on image data only, and augment this data set using generative artificial intelligence to maintain relevant features, thereby expanding the training data set and improving model performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If training data is collected from limited participant studies in vehicles with extensive sensor technology, then the quality and ground truth of training data is improved, but the quantity of training data remains insufficient
Solution Approach 1:
The patent creates synthetic copies of real training data by training a generative AI model on a small set of real multimodal training data, then using this model to generate additional synthetic training data that preserves the statistical properties and ground truth labels of the original data
Solution Approach 2:
The patent transforms the training data from multimodal (multiple sensor types) to unimodal (image-only) by selectively retaining only image data while preserving ground truth labels, thereby changing the data parameters to match available sensor configurations in production vehicles
2Adaptability or versatility
If a multimodal training data set is reduced to image data only, then the adaptability to production vehicles with limited sensors is improved, but the information content and model performance may deteriorate
Solution Approach 1:
The patent uses generative AI to create synthetic image data that replicates the statistical properties and information content of the original multimodal data, effectively copying the essential information patterns without requiring the actual sensor hardware
Solution Approach 2:
The patent extracts only the image data modality from the multimodal training set while retaining the ground truth labels, separating the essential visual information from the unavailable sensor data to create a standalone training solution
Data Source
Figure 1~2
Figure 1
Figure 3
AI summary
The invention relates to a method for providing a machine learning model (20) for motion sickness detection in a vehicle (50), wherein a multimodal training dataset (10) with a motion sickness state (70) as the basic truth (14) is obtained, wherein a training dataset (11) is generated from the multimodal training dataset (10) in which only image data (12) are considered as the modality, wherein the generated training dataset (11) is augmented, wherein features (16) to be retained in the image data (12) are defined for this purpose, and wherein further training data are generated by generating further image data (15) using an image-generating generative artificial intelligence (17) while retaining the defined features (16), wherein a machine learning model (20) is trained to estimate a motion sickness state (70) of an occupant based on image data (12, 15).wherein the machine learning model (20) is trained using at least a part of the augmented training dataset (18), and wherein the trained machine learning model (20) is provided. Furthermore, the invention relates to a device (1) for providing a machine learning model (20) for motion sickness detection. Methods and a device (52) for detecting motion sickness in a vehicle occupant (50) are also described.