Adaptive Red-Eye Detection Using Dynamic Template Updates
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing red-eye detection systems face challenges in accurately detecting red eyes due to variations in eye shape caused by different degrees of eye opening, and using multiple templates significantly reduces processing speed.
Innovation Solution
A red-eye detection device that generates and updates red-eye templates corresponding to varying degrees of eye opening, using eye opening degree calculation and relative eye opening degree calculation to normalize eye opening representation, and optionally incorporates drowsiness estimation and correlation learning with black eye size to improve accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If template matching is performed using a single red-eye template, then processing speed is maintained, but red-eye detection accuracy deteriorates when eye opening degree varies
Solution Approach 1:
The patent applies dynamics by making the red-eye template adaptive to the detected eye opening degree. Instead of using a fixed single template, the system dynamically selects or generates appropriate templates based on the current eye opening state, allowing the detection system to adapt its parameters to varying conditions while maintaining efficient processing through automated selection.
Solution Approach 2:
The patent changes the template parameters based on eye opening degree detection. By adjusting template characteristics (such as size, shape, or intensity distribution) according to the detected eye opening state, the system optimizes detection accuracy for different eye conditions without requiring manual intervention or complex multiple-template processing.
2Measurement precision
If template matching is performed using a plurality of red-eye templates corresponding to all degrees of eye opening, then red-eye detection accuracy improves, but processing speed significantly deteriorates
Solution Approach 1:
The patent segments the eye opening degree range into distinct states (e.g., closed, partially open, fully open) and assigns specific templates to each segment. This segmentation allows the system to use a limited number of specialized templates rather than processing all possible variations, maintaining high detection accuracy for each eye state while keeping processing speed acceptable through reduced template comparison overhead.
Solution Approach 2:
The patent performs preliminary classification of eye opening degree before template matching. By first detecting and categorizing the eye opening state, the system pre-selects the appropriate template subset, avoiding unnecessary comparisons with all possible templates. This preliminary action significantly reduces processing time while ensuring the correct template is used for accurate detection.
3Device complexity
If distance from upper eyelid to lower eyelid is used as eye opening measure, then calculation is simple, but detection reliability deteriorates due to variations with camera distance and individual differences
Solution Approach 1:
The patent introduces relative eye opening degree as an intermediary metric that normalizes the raw pixel distance measurement. By establishing a reference scale (such as the distance between eye corners or face width) and expressing the eyelid distance relative to this reference, the system eliminates the need for absolute measurements, thereby compensating for variations in camera distance and individual anatomy while maintaining calculation simplicity.
Data Source
AI summary
An ECU connected to an image sensor includes a face position and face feature point detection unit that detects the feature points of the face of the driver, a red-eye detection unit that detects the red eye with template matching using a red-eye template, an eye opening degree calculation unit that calculates the degree of eye opening, a relative eye opening degree calculation unit that calculates the relative degree of eye opening which is 0% in an eye-closed state and is 100% in an eye-open state, and a red-eye template update unit that generates a red-eye template on the basis of the relative degree of eye opening and updates a red-eye template used for the next template matching with the generated red-eye template.