Potential Obstacle Detection Using RGB to Depth Image Encoder-Decoder Network: Application to Unmanned Aerial...

Area of Science:

• Robotics and Automation
• Computer Vision
• Artificial Intelligence

Background:

• Unmanned Aerial Vehicles (UAVs) require robust real-time object detection for safe navigation.
• Existing methods for depth estimation from single RGB cameras often face challenges in balancing accuracy and computational complexity.
• Accurate distance estimation is crucial for identifying potential collision sources.

Purpose of the Study:

• To propose a novel method for real-time object detection and distance estimation for UAVs using a single RGB camera.
• To develop an efficient encoder-decoder network for RGB-to-depth mapping.
• To ensure the developed system achieves a practical trade-off between computational complexity and detection accuracy.

Main Methods:

• Development of a new encoder-decoder network architecture for RGB-to-depth mapping.
• Implementation of a specialized algorithm to refine distance predictions and compensate for measurement inaccuracies.
• Real-time testing and validation on a micro-drone equipped with a front-facing RGB camera in an indoor environment.

Main Results:

• The proposed network achieved efficient real-time performance with only 6.3 million parameters, comparable to models with significantly more parameters.
• Demonstrated a satisfactory balance between network complexity and accuracy in RGB-to-depth mapping compared to existing methods.
• Successful practical implementation and testing of the complete collision detection solution on a micro-drone.

Conclusions:

• The developed single RGB camera-based system provides an efficient and accurate solution for real-time collision avoidance in UAVs.
• The method offers a practical approach for enhancing UAV safety through rapid distance estimation.
• Availability of all data, source codes, and pre-trained weights facilitates reproducibility and practical deployment.