Eye State Detection Using AdaBoost Genetic Programming
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing vision-based driver monitoring systems face challenges in accurately determining the open vs. closed eye state of a driver using classifiers like neural networks, which are complex and resource-intensive, limiting their use in cost-sensitive applications.
Innovation Solution
A method combining genetic programming and adaptive boosting to develop classification algorithms, where a genetic programming loop is embedded within an adaptive boosting loop to create and weight classifiers for determining the open vs. closed eye state from video image data, reducing computational requirements while maintaining high accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If neural network or support vector machine classifiers are used to achieve high classification accuracy, then classification accuracy is improved, but device complexity and processing capability requirements increase
Solution Approach 1:
The patent segments the classification task into multiple stages: first using a simple AdaBoost classifier with basic features to make a preliminary classification, then only proceeding to more complex analysis if needed. This segmentation allows the system to achieve high accuracy while maintaining low average complexity by avoiding complex classifiers for most cases.
Solution Approach 2:
The patent introduces an intermediary approach by using AdaBoost as a bridge between simple rule-based systems and complex neural networks. AdaBoost provides a middle-ground solution that achieves comparable accuracy to neural networks but with significantly reduced computational requirements and simpler implementation.
2Measurement precision
If neural network or support vector machine classifiers are used to achieve high classification accuracy, then classification accuracy is improved, but processing capability and memory requirements increase
Solution Approach 1:
The patent employs lightweight, computationally inexpensive classifiers (AdaBoost with simple features) that consume minimal processing power and memory. These 'cheap' classifiers are designed to be disposed of or replaced easily, providing a cost-effective solution that achieves sufficient accuracy without the heavy resource demands of neural networks.
Solution Approach 2:
The patent changes the parameters of the classification approach by using a different algorithm (AdaBoost instead of neural networks), adjusting feature selection to use simpler eye aspect ratio and blink detection metrics, and modifying the decision threshold to optimize for low-power operation while maintaining acceptable accuracy levels.
3Measurement precision
If complex classifiers are used to accurately determine eye state, then measurement precision is improved, but ease of manufacture and implementation deteriorates
Solution Approach 1:
The patent implements a self-service classification system where the AdaBoost classifier automatically adapts to different drivers and conditions without requiring complex manual configuration or training. The system self-adjusts by learning from incoming data and making real-time decisions, eliminating the need for complex setup procedures or expert intervention.
Data Source
Figure 1
Figure 2
AI summary
Real-time video images of a human subject's face are processed by a plurality of classification algorithms (h1 ... hT) developed in an off-line training process to determine the open vs. closed eye state (S) of the subject. The off-line training process utilizes a genetic programming loop (18) embedded within an adaptive boosting loop (20-22), and forms the classification algorithms and weighting factors for combining their classification scores. In the real-time process, the individual classification scores are combined and compared to a threshold to determine the open vs. closed eye state.