Assistive app based on eye tracking

Student’s Name:Yu Qiu

Student's ID: 2036484
Supervisor: Arodh Lal Karn

The objective of this thesis is to explore the feasibility and effectiveness of implementing a
real-time eye-tracking system using the YOLOv5 object detection model, optimized for mobile
platforms through TensorFlow Lite. This research addresses the challenge of deploying highperformance deep learning models on resource-constrained devices, such as smartphones, by
utilizing model compression techniques within TFLite to reduce computational demand without
substantially sacrificing accuracy.
The implementation of YOLOv5, adapted for eye-tracking, demonstrates the model’s versatility
and robustness in detecting eye positions under various environmental conditions. The system
leverages TensorFlow Lite’s capabilities for model quantization and optimization to ensure that
the application runs efficiently in real-time on Android devices. The results indicate that the
optimized YOLOv5 model achieves satisfactory accuracy and latency on standard mobile hardware,
making it suitable for a wide range of applications from assistive technologies to user
behavior analysis.
This study not only contributes a practical solution to mobile-based eye-tracking but also provides
a foundation for future research in enhancing the accessibility and usability of eye-tracking
technologies through advanced machine learning techniques. Further explorations into adaptive
quantization and dynamic model adjustments could offer improvements in system performance
and energy efficiency, broadening the potential use cases for mobile-based eye-tracking solutions.

The integration of YOLOv5 with TensorFlow Lite represents a significant enhancement in the
field of mobile vision applications, specifically in creating an accessible, efficient eye-tracking system.
This research not only confirmed the viability of using YOLOv5 within resource-constrained
environments but also showcased how TFLite’s model compression and optimization techniques
can be utilized to enhance performance without compromising accuracy significantly. The system
developed through this study provides a practical tool for real-time eye-tracking, offering
potentials for application in user interface accessibility for the disabled, behavioral analysis,
and in interactive gaming. Future directions for this research include exploring further efficiency
improvements through quantization and pruning techniques in TFLite, and expanding
the system’s capabilities to include predictive eye movement analytics, which could revolutionize
user-device interactions.uture work will aim to further optimize the detection process and
expand the application areas of this technology.