Overview
In aviation safety, managing Foreign Object Debris (FOD) on air stations is a pivotal challenge. Traditional FOD detection methods such as mechanical sweepers often fall short in efficiency. In partnership with Skylark Labs and the Indian Navy, an innovative FOD Detection System was developed to revolutionize FOD management and enhance the safety and operational readiness of aircraft.
Methods
- Explored various architecture configurations integrating SOTA super-resolution models with detection and classification models
- Experimented with different dataset compositions, segmenting FODs into different class counts with a 'non-FOD' class to reduce false positives
- Created a synthetic dataset exceeding 80 GB with around 40,000 images covering diverse FOD types
- Ensured translational invariance and diverse lighting conditions: sunlight, overcast, rain, shadows, and water reflections
Outcomes
- YOLOv8 and Vision Transformer models both demonstrated high FOD detection accuracy
- Vision Transformer showed particular proficiency in ignoring irrelevant objects, reducing false positives
- System significantly increased detection speed, enhancing air station operational efficiency
- Diverse synthetic datasets ensured model effectiveness across real-world environmental conditions
My Contributions
- Explored and tested various model architecture combinations with super-resolution integration
- Developed a synthetic dataset generation pipeline simulating various lighting and weather conditions
- Built a translational invariant dataset pipeline for model robustness to FOD position variation
- Contributed to model robustness improvements that enabled real-world deployment readiness