Ground Truth Data Generator for Eye Location on Infrared Driver Recordings.

Area of Science:

• Computer Vision
• Machine Learning
• Automotive Engineering

Background:

• Manual image labeling for training neural networks is expensive and time-consuming.
• High costs associated with image labeling significantly impact automotive driver monitoring projects.
• Generating accurate ground truth data is crucial for developing robust AI systems.

Purpose of the Study:

• To present an algorithm for automatically generating ground truth data for 2D eye location in infrared driver images.
• To reduce the cost and time associated with creating datasets for driver monitoring systems.
• To enable the training of neural networks with high accuracy and consistency for eye detection.

Main Methods:

• Development of an algorithm for 2D eye location detection in infrared images.
• Implementation of detection restrictions to enhance algorithm accuracy.
• Generation of a dataset for training neural networks without human modification.

Main Results:

• The algorithm produces highly accurate, though not always consistent, ground truth data for 2D eye location.
• The generated dataset is intended for training neural networks to achieve superior eye detection accuracy and consistency.
• Demonstration of the feasibility of automatically generating high-quality ground truth data for driver monitoring.

Conclusions:

• Automated generation of high-quality ground truth data for neural network training is achievable.
• This approach addresses a significant challenge in the automotive industry regarding data generation for driver monitoring.
• The developed algorithm offers a cost-effective and efficient solution for creating essential datasets.