Reproduce “Autonomous FPV Flight with Translational Optical Flow and Uncertainty Mask” on OpenResearch

About this paper

Yang Deng, Yu Hu, Feng Yu, Linzuo Zhang, Danping Zou

Researchers from Shanghai Jiao Tong University developed an autonomous FPV quadrotor flight system utilizing a monocular camera, which processes visual input through translational optical flow and an uncertainty mask. This method enabled robust, agile flight at real-world speeds up to 11.79 m/s with a 93.3% success rate, significantly exceeding prior benchmarks for monocular-vision-based UAV obstacle avoidance.

Key insights

• Disentangling translational and rotational optical flow components significantly improves a learning-based policy's ability to interpret scene structure and navigate obstacles during aggressive maneuvers.
• Complementary visual cues, such as an uncertainty mask derived from flow consistency, are crucial for robust obstacle avoidance in regions where raw optical flow is inherently unreliable, like near the Focus of Expansion.
• Integrating classical vision concepts like flow decomposition and consistency checks into a task-oriented learning framework for control leads to more robust and generalizable monocular navigation systems.

Summary by alphaXiv.