Gaze Detection Using 3D Face Model Coordinate Conversion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing gaze detection methods using single-lens cameras struggle to accurately estimate gaze direction without prior calibration or learning of pupil patterns, especially when the subject's head is at an arbitrary posture.
Innovation Solution
A gaze detection apparatus that extracts feature points from an input image, including a pupil, and uses a three-dimensional face model to convert and estimate the gaze direction, eliminating the need for calibration or prior learning of pupil patterns, by employing a feature point detection unit, a three-dimensional face model storage unit, and a gaze estimating unit.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a single-lens camera is used for gaze detection, then the system complexity is reduced, but the gaze direction cannot be accurately detected without prior calibration or learning
Solution Approach 1:
The patent pre-stores multiple three-dimensional face models representing different head postures and corresponding pupil position information. This preliminary preparation allows the system to directly match and estimate gaze direction without requiring calibration or learning procedures during actual operation, thus maintaining low system complexity while achieving accurate measurement.
Solution Approach 2:
The patent creates three-dimensional virtual copies of face models with known geometric relationships between facial features and pupil positions. These copied models serve as reference templates that can be directly compared with actual images to determine gaze direction, eliminating the need for complex calibration procedures.
2Ease of operation
If the head posture is fixed, then the gaze detection can be performed with simpler methods, but the system cannot handle arbitrary head postures
Solution Approach 1:
The patent segments the continuous space of head postures into multiple discrete representative postures, each with its own three-dimensional face model. This segmentation allows the system to handle various head positions using simple, pre-prepared models rather than requiring complex continuous adaptation mechanisms.
Solution Approach 2:
The patent creates a universal set of three-dimensional face models that can handle multiple head postures simultaneously. These models serve multiple functions: representing different postures, providing reference pupil positions, and enabling gaze estimation across all postures, thus achieving versatility without sacrificing operational simplicity.
3Measurement precision
If pattern dictionaries by pupil directions are prepared in advance, then the gaze detection accuracy is improved, but the learning process becomes time-consuming and complex
Solution Approach 1:
The patent pre-calculates and stores the correspondence between head posture parameters and pupil positions in three-dimensional space. This preliminary computation of geometric relationships eliminates the need for time-consuming learning during actual use, as the system can directly apply these pre-established relationships for accurate gaze detection.
Data Source
AI summary
An image input unit, a feature point detection unit configured to extract at least four image feature points including a feature point of a pupil and which do not exist on an identical plane from an input image, a three-dimensional face model storage unit configured to store shape information of a three-dimensional face model and at least coordinates of reference feature points on the three-dimensional face model corresponding to the feature points extracted by the feature point detection unit, a converting unit configured to convert a coordinate of the feature point of the pupil onto surface of the three-dimensional face model on the basis of the correspondence between the extracted feature points and the reference feature points, and a gaze estimating unit configured to estimate the gaze direction from the converted coordinate of the pupil are provided.