Neural Network Gaze Direction Estimation Without Head Pose Calibration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current gaze direction estimation systems are unreliable due to errors in deriving necessary user information, such as head pose and user dependency, leading to incorrect determinations.
Innovation Solution
A method using a neural network to detect the ellipse representing the iris in image data, outputting two possible gaze direction vectors with associated probabilities, allowing determination of the gaze direction without requiring head pose or user-specific calibration.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If current gaze direction estimation systems are used, then gaze direction can be determined, but the system reliability deteriorates due to multiple error sources including missing prior images, erroneous information, and false outputs from underlying modules
Solution Approach 1:
The patent extracts and eliminates the dependency on error-prone external modules (head tracker, user authentication, prior eye images) by implementing a self-contained neural network system that processes only the current eye image to directly determine gaze direction, thereby improving reliability without proportionally increasing complexity
Solution Approach 2:
The patent uses a pre-trained neural network model that has been trained on large datasets to replicate human gaze direction judgment, replacing complex multi-module systems with a single trained model that can reliably determine gaze from eye images alone
2Measurement precision
If user-dependent calibration and prior user information are required, then gaze direction estimation can be performed, but the loss of time increases due to calibration procedures and the need for prior images
Solution Approach 1:
The neural network system performs self-calibration through pre-training on diverse datasets, eliminating the need for user-specific calibration procedures. The model adapts to different users and conditions during training, enabling immediate accurate gaze estimation without time-consuming setup
Solution Approach 2:
The neural network is pre-trained on extensive datasets containing various eye images, lighting conditions, and gaze directions before deployment. This preliminary training action ensures the model is ready to accurately estimate gaze direction for any user immediately, without requiring calibration time
3Reliability
If multiple sources of information (head pose, biometric features, prior images) are used, then gaze direction can be determined, but the device complexity increases due to multiple required modules
Solution Approach 1:
The neural network is designed as a universal system that can determine gaze direction for any user without requiring separate modules for head tracking, user authentication, or prior image storage. The single model handles all these functions through its training, reducing system complexity while maintaining reliability
Solution Approach 2:
The patent removes the dependency on external error-prone modules by extracting the essential function (gaze determination) and implementing it directly within the neural network through image processing, eliminating the need for separate head trackers, authentication systems, and prior image databases
4Measurement precision
If conventional gaze estimation systems are used, then gaze direction can be determined, but the productivity decreases due to the need for multiple images and complex processing steps
Solution Approach 1:
The neural network skips intermediate processing steps (head pose estimation, biometric feature extraction, prior image comparison) and directly processes the eye image to determine gaze direction in a single operation, significantly improving processing speed while maintaining accuracy through the model's training
Data Source
Figure 1~2
Figure 3
Figure 4
AI summary
Disclosed is a method for determining a gaze direction of a user (5) by employing a neural network (220) operating on the basis of image data of an image of a face of the user (5). The neural network (220) detects an ellipse (20) representing an outer border of an iris (17) of the user's eye (15) and outputs a first vector (31), a second vector (32) and a respective first and second probability (pi, pz) representing the probability that the first vector (31) or the second vector (32) is the gaze direction Further disclosed are an apparatus configured to perform such method and a vehicle comprising such apparatus.