Sight Tracking Using ELM Neural Network and Kalman Filter
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing sight tracking methods based on kalman filtering have lower precision and speed in determining the position of fixation points of human eyes on a screen due to the linear state equation and instability in prediction regions.
Innovation Solution
A sight tracking method that uses an Extreme Learning Machine (ELM) neural network to determine a target model from visual feature parameters, which modifies the prediction region obtained by kalman filtering to improve accuracy and speed in determining the fixation point, by adjusting the kalman gain equation with an adjustment factor.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If kalman filtering method with linear state equation is used to track sight, then the interaction procedure can be completed, but the accuracy and speed of determining fixation point position are lower
Solution Approach 1:
The patent changes the mathematical model from a linear state equation to a nonlinear state equation that incorporates eye movement characteristics. Specifically, it uses a state equation that accounts for the rotational movement of the eyeball and the geometric relationship between the iris center and fixation point, thereby improving measurement precision without requiring complete system redesign
Solution Approach 2:
The patent replaces the traditional mechanical/optical tracking approach with a computational approach using kalman filtering combined with neural network prediction. This substitution allows for more accurate modeling of eye movement dynamics while maintaining system simplicity through software-based solutions
2Measurement precision
If kalman filtering method is used to predict iris center location, then the tracking process can proceed, but the prediction accuracy is lower due to linear equation limitations
Solution Approach 1:
The patent performs preliminary action by pre-establishing the nonlinear state equation model that captures eye movement patterns. The model is prepared in advance with predefined parameters for eyeball rotation and iris movement, allowing for faster real-time prediction without requiring complex calculations during the actual tracking process
Solution Approach 2:
The patent creates a computational model that copies and simulates the physical eye movement process. By modeling the relationship between iris center position and fixation point location through mathematical equations, it achieves accurate prediction without requiring direct measurement of the fixation point, thus reducing calculation time
3Productivity
If traditional kalman filtering is used for sight tracking, then the system can operate, but the precision and speed of determining fixation point position are reduced
Solution Approach 1:
The patent introduces dynamics by using a state equation that adapts to the changing position of the iris center. The model dynamically updates the prediction based on the current iris position and the geometric relationship to the fixation point, allowing the system to maintain high precision and speed even as the eye moves across different regions of the screen
Data Source
AI summary
Embodiments of the present disclosure relate to a sight tracking method and a device, the sight tracking method comprises: determining an observation region where an iris center of a to-be-tested iris image is located according to a target model; modifying a prediction region by using the observation region, to obtain a target region, the prediction region being a region where the iris center of the to-be-tested iris image is located determined by a kalman filtering method; and determining a position of fixation point of human eyes on a screen according to the target region.