View Code
2025.04 - 2025.05 Team Project Computer Vision / AI

FALCON: Airport Safety AI

A comprehensive airport safety system detecting Bird Strikes, FOD (Foreign Object Debris), and ground incursions. Leverages Synthetic Data from Unity/Blender for robust AI training.

Repository Back to Projects

Project Overview

FALCON enhances airport safety by providing real-time risk discovery and decision support. It consists of "Hawkeye" for ground and tower control monitoring, and "RedWing" for pilot assistance.

Key Engineering Contributions

Synthetic Data Pipeline

Built an end-to-end data generation pipeline using Unity and Blender scripting. This automated system created thousands of labeled images for rare "long-tail" events like bird strikes and runway debris, solving the critical lack of real-world training data.

Synthetic Data Generation

Pilot Assist AI (RedWing)

Implemented the core logic for the "RedWing" system, integrating YOLOv8 for hazard detection and a Bird Strike Risk Analysis (BDS) model. Optimized inference for real-time alerts to pilots.

Bird Detection System

Action Recognition

Designed and trained a Temporal Convolutional Network (TCN) to interpret complex marshaller hand signals. This transformed continuous pose estimation data into actionable commands for the digital cockpit system.

Pose Estimation

Core Technology

  • Object Detection: YOLOv8 pipelines trained on hybrid (real + synthetic) datasets.
  • Action Recognition: TCN (Temporal Convolutional Network) for interpreting complex marshaller hand gestures (98.99% accuracy).
  • Simulation: Custom Unity-based airport simulator (RunwaySim) for generating training scenarios.
  • Pose Estimation: YOLOv8n-pose for detecting falls and worker status.

Tech Stack

YOLOv8 / YOLOv11 Unity 3D Blender Python (PyTorch) PyQt Flask MySQL

Technical Challenges & Solutions

Data Scarcity for Rare Events

Problem: Lack of real-world footage for bird strikes and specific airport accidents.
Solution: Built a synthetic data pipeline using Blender and Unity to generate thousands of photorealistic, automatically labeled training images.

Pose Estimation for Falls

Problem: Standard models struggled to detect people lying down or in unusual orientations.
Solution: Generated synthetic pose data with various camera angles and body orientations to fine-tune the model.