Flight safety boosted with AI-powered obstacle detection
The outcomes
- Obstacle detection model is built**,** starting with 12 synthetic scenes and 2 real scenes
- Model training can be sped up with the use of cost-effective synthetic data.
The context
A global leader in helicopter manufacturing aimed to improve safety for civilian helicopters by addressing a critical challenge: detecting thin obstacles during low-altitude flights. These objects—wires, poles, fences—are often invisible to standard flight systems, posing serious risks in missions like mountain rescues. Indeed, due to their limited maneuverability, helicopters find it challenging to avoid these obstacles quickly, increasing the likelihood of accidents.
The logic
The helicopter manufacturer needed a cost-effective, accurate detection system that could work in varied environments without relying on expensive radar technology used in military helicopters.
The solution
We built a computer vision-based depth estimation model that monitors camera images and alerts pilots when obstacles approach. The biggest hurdle was data availability: training required diverse scenarios to avoid overfitting.
- Real-world data was captured using drones for authentic conditions.
- Synthetic data was generated to simulate rare or extreme environments, such as snowstorms or rugged terrain, ensuring comprehensive coverage.
Synthetic data proved a game-changer—enabling large-scale, customizable datasets that accelerated development and improved model robustness. It ensured a wide range of examples for training the model.