Unmanned aerial vehicle end-to-end safe autonomous obstacle avoidance navigation method based on monocular vision depth estimation
By combining the Depth Anything V2 model with an actor-critic reinforcement learning strategy, we achieved end-to-end safe and autonomous obstacle avoidance navigation for UAVs. This solved the ill-conditioned optimization and environmental adaptability problems of monocular visual obstacle avoidance in low-altitude, high-speed flight, and achieved real-time, highly robust obstacle avoidance performance.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- DIFFERENTIAL ZHIFEI (HANGZHOU) TECHNOLOGY CO LTD
- Filing Date
- 2026-04-20
- Publication Date
- 2026-07-17
AI Technical Summary
Existing monocular vision obstacle avoidance methods suffer from ill-conditioned optimization problems caused by texture loss, motion blur, and drastic changes in viewpoint during low-altitude, high-speed flight of UAVs. Furthermore, deep learning models exhibit significant biases when the environment changes, failing to guarantee safety and real-time performance.
The Depth Anything V2 monocular depth estimation model is seamlessly coupled with an actor-critic reinforcement learning strategy. By using a pre-trained deep learning network, geometric priors can be quickly obtained, enabling an end-to-end perception-decision-control closed loop and reducing computational complexity and inference latency.
Real-time, robust monocular visual obstacle avoidance was achieved on a resource-constrained UAV platform, improving obstacle avoidance adaptability and system compactness in complex environments.
Smart Images
Figure CN122408773A_ABST