Eye Tracking Model Training Using Reference Gaze Data

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Traditional methods for training machine learning-based eye tracking algorithms require large amounts of annotated data, which is time-consuming and costly to collect, and may not capture real-life behavior effectively.

Innovation Solution

A method for training an eye tracking model using sensor data from a first eye tracking sensor and reference eye tracking data from a second eye tracking sensor, allowing the model to predict eye tracking data and improve its accuracy over time.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Quantity of substance

If traditional methods are used to collect annotated training data, then the training data can be obtained, but it requires large amounts of time and resources to collect and annotate

Engineering Contradiction:

Improveamount of training dataVSAvoidtime to collect and annotate data

Core Design Contradiction:

Quantity of substanceVSLoss of time

Solution Approach 1:

The patent uses an existing calibrated eye tracker to generate reference gaze data that serves as training labels for the machine learning model. Instead of manually annotating training data, the system copies the functional capability of the calibrated eye tracker through software-based gaze estimation from images captured by the uncalibrated eye tracker, significantly reducing annotation time and costs

Inventive Principle:

Principle #26Copying

Solution Approach 2:

The system performs self-calibration by using the uncalibrated eye tracker itself to collect training data and automatically train the machine learning model. The eye tracker captures images during normal use, and the ML model is trained on these images with gaze points derived from the reference eye tracker, allowing the system to improve its own performance without external intervention

Inventive Principle:

Principle #25Self-service

2Reliability

If traditional calibration techniques are used, then the eye tracking system can be set up, but it requires visible calibration targets and a calibration process that interrupts normal use

Engineering Contradiction:

Improveaccuracy of eye trackingVSAvoidconvenience of setup

Core Design Contradiction:

ReliabilityVSEase of operation

Solution Approach 1:

The patent performs calibration data collection during normal system operation rather than requiring a separate calibration session. The uncalibrated eye tracker continuously captures images, and the ML model is trained in the background using these images and reference gaze data, so calibration occurs preliminarily and continuously without interrupting the user's normal interaction with the system

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The patent replaces the mechanical calibration process that requires visible targets and manual adjustment with a software-based solution. The machine learning model substitutes for the traditional calibration algorithm, using image processing and pattern recognition to estimate gaze positions without requiring the user to follow calibration procedures or look at specific calibration targets

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

3Measurement precision

If machine learning is used to train an eye tracking algorithm, then the system can perform eye tracking, but it requires plenty of training data that may not reflect real-life scenarios

Engineering Contradiction:

Improveaccuracy of gaze estimationVSAvoidability to handle diverse scenarios

Core Design Contradiction:

Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The patent enables continuous collection of training data during normal system operation. The uncalibrated eye tracker captures images throughout the day, and the ML model is continuously trained on this diverse real-life data, ensuring the training data reflects actual usage scenarios rather than controlled calibration conditions. This continuous data collection and training process improves both accuracy and adaptability to various real-world situations

Inventive Principle:

Principle #20Continuity of useful action

Solution Approach 2:

The system dynamically adapts to different users and scenarios by continuously updating the ML model with new training data. The calibration process is not static but evolves over time as the system collects more diverse images during normal use, allowing the model to adapt to individual user characteristics and various real-life scenarios encountered during operation

Inventive Principle:

Principle #15Dynamics

Data Source

PatentEP4083755B1Training an eye tracking model

Publication Date: 2025.02.19 TOBII TECH AB

Figure 1~2
Figure 3
Figure 4

AI summary

A method (300) for training an eye tracking model (710) is disclosed, as well as a corresponding system (400) and storage medium. The eye tracking model is adapted to predict eye tracking data based on sensor data (709) from a first eye tracking sensor (411). The method comprises receiving (301) sensor data obtained by the first eye tracking sensor at a time instance and receiving (302) reference eye tracking data for the time instance generated by an eye tracking system (420) comprising a second eye tracking sensor (421). The reference eye tracking data is generated by the eye tracking system based on sensor data obtained by the second eye tracking sensor at the time instance. The method comprises training (303) the eye tracking model based on the sensor data obtained by the first eye tracking sensor at the time instance and the generated reference eye tracking data.