Eye Gaze Tracking Using Camera Segmentation and Prediction Models

The patent segments the eye gaze tracking task into multiple independent modules: eye detection module for locating eyes, gaze direction calculation module for computing gaze direction, and position mapping module for converting to screen coordinates. This segmentation allows each module to be optimized independently, improving overall accuracy while maintaining manageable system complexity through modular architecture.

The patent introduces an intermediary coordinate transformation process that bridges the gap between eye coordinate systems and display screen coordinate systems. This intermediary mapping layer, combined with the prediction mechanism, enables accurate gaze position determination without requiring direct complex measurements, thus improving precision while managing system complexity.

The patent applies preliminary action by using a prediction mechanism that anticipates gaze position based on eye motion trends. The system predicts the next gaze position before the user actually looks at it, reducing the real-time computational burden and improving response speed. The prediction model processes historical eye position data to forecast future gaze locations, enabling faster response without excessive computational load.

The patent implements dynamic adaptation by adjusting the tracking methodology based on real-time conditions. When eye motion is detected, the system switches to prediction mode for faster response; when eye motion is minimal, it uses direct detection mode. This dynamic switching optimizes the balance between response speed and computational load, allowing the system to respond quickly without consistently consuming high computational resources.

The patent segments the deep learning processing into two parts: an offline training phase where the prediction model is trained using deep learning on historical data, and an online execution phase where the pre-trained model makes rapid predictions. This segmentation allows the computationally intensive deep learning to be performed once during training, improving accuracy without significantly increasing real-time processing time during actual eye gaze tracking.

The patent applies preliminary action by pre-training the prediction model using deep learning algorithms on extensive historical eye gaze data before actual tracking begins. During real-time operation, the pre-trained model makes fast predictions without requiring continuous deep learning computation. This preliminary training phase captures complex patterns and relationships, enabling accurate gaze position determination with reduced real-time processing requirements.