An Unmanned Aerial Vehicle Indoor Low-Computation Navigation Method Based on Vision and Deep Learning

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Indoor unmanned aerial vehicle (UAV) applications are increasing, often requiring consistent navigation paths.
• Existing indoor positioning systems can be costly and computationally intensive, exceeding application needs.
• There is a need for efficient, low-cost solutions for indoor UAV path following and obstacle avoidance.

Purpose of the Study:

• To develop a computationally efficient and cost-effective solution for indoor UAV path following and obstacle avoidance.
• To enhance UAV navigation capabilities in environments with consistent paths and potential obstacles.

Main Methods:

• Utilized a down-looking camera for path following, refining the carrot casing algorithm and introducing a novel Line-Fitting Path-Following (LFPF) algorithm.
• Employed a front-looking camera with depth images and YOLOv4-tiny for real-time obstacle detection and avoidance strategy implementation.
• Tested the system on an Nvidia Jetson Nano, an entry-level computing platform.

Main Results:

• The LFPF algorithm demonstrated superior performance in adapting to light variations and maintaining consistent flight speed compared to the carrot casing algorithm.
• The LFPF algorithm maintained an error margin within ±40 cm in real-world indoor flight scenarios.
• The system achieved minimal computational demands, running at 23 FPS on the Nvidia Jetson Nano.

Conclusions:

• The proposed system offers a cost-effective and computationally efficient solution for indoor UAV navigation.
• The LFPF algorithm provides robust path following under varying light conditions and dynamic speeds.
• The integrated obstacle avoidance system effectively manages threats based on type and proximity, ensuring safe operation.