Eye Detection Using Weighted Candidate Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing eye monitoring systems in vehicles face challenges in accurately detecting eyes due to variations in light and head rotation, often resulting in false detections and requiring subject calibration and full visibility of the driver's face.
Innovation Solution
A video imaging system that uses a camera to capture images of a subject's face, processing them with a processor that detects potential eye candidates based on appearance, spatial-geometrical constraints, and physiological skin properties, assigning weighted values to filter and select the most likely eye candidates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional eye detection methods are used, then the system can detect eyes in controlled conditions, but the detection accuracy deteriorates under variations in light and head rotation
Solution Approach 1:
The detection process is segmented into multiple independent filtering stages: initial candidate detection, appearance-based filtering, spatial-geometrical constraint filtering, and physiological skin property filtering. Each stage processes specific features independently and combines results through weighted voting, allowing the system to handle variations in light and head rotation by evaluating multiple independent criteria.
Solution Approach 2:
The system changes multiple parameters simultaneously including appearance features (texture, color), spatial-geometrical constraints (position, size, orientation), and physiological skin properties (reflectance, texture). By evaluating multiple parameters and their combinations, the system maintains detection accuracy across varying lighting conditions and head rotations.
2Reliability
If multiple filtering criteria are applied to improve detection accuracy, then the reliability of eye detection improves, but the computational complexity and processing time increase
Solution Approach 1:
The system performs preliminary filtering by detecting potential eye candidates first, then applies subsequent filtering stages only to these candidates. This preliminary action reduces the number of regions requiring detailed analysis, maintaining reliability through multiple filters while reducing overall computational complexity by processing only relevant regions.
Solution Approach 2:
The system applies multiple filtering criteria (appearance, spatial-geometrical, physiological) that may be more than strictly necessary, but uses weighted voting to combine them efficiently. This partial application of excessive filters ensures high reliability while the weighted combination mechanism keeps processing manageable by prioritizing the most informative features.
3Measurement precision
If the system requires full visibility of the driver's face for accurate eye detection, then detection accuracy improves, but the ease of operation deteriorates as it becomes more restrictive
Solution Approach 1:
The system focuses computational resources on local regions identified as potential eye candidates rather than requiring analysis of the entire face. By detecting candidates based on local appearance features and then applying filters only to these localized regions, the system maintains high detection accuracy while operating effectively even when full face visibility is not available.
Data Source
Figure 1
Figure 2
Figure 3~4
AI summary
A system (30) and method (100) are provided for detecting one or both eyes (22) of the driver (20) of a vehicle (10). The system (30) includes a video imaging camera (24) oriented to generate images of the driver's face. The system (30) also includes a video processor (32) for processing the images generated with the video imaging camera (24). Filters are applied to each of potential eye candidates (50A-50F) to determine which candidates represent an eye (22), weighted values are assigned to the filtered eye candidates, and an eye (22) is detected based on the weighting of each eye candidate. According to one aspect, different size patches of potential eye candidates are processed and compared to models.